Duplicate from xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16

Co-authored-by: Florian <[email protected]>

.gitattributes

@@ -0,0 +1,59 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-12B-v2/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2/acc-vs-budget.png filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2/accuracy_chart.png filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-BF16/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-EA/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-NGC/acc-vs-budget.png filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-NGC/accuracy_chart.png filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-NGC/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-rc0807/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2-rc0807-safetensors/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+NVIDIA-Nemotron-Nano-9B-v2.bak/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nemotron-H-4B-Base-8K/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nemotron-H-4B-Instruct-128K/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nemotron-H-8B-Base-8K/nemo/context/nemo_tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nemotron-H-8B-Base-8K/nemo/weights/.metadata filter=lfs diff=lfs merge=lfs -text
+Nemotron-H-8B-Base-8K/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nemotron-Nano-3-30B-A3.5B-dev-1016/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Nvidia-Nemotron-Nano-v2-9B-0805/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Qwen3-14B/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Qwen3-8B/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+llama-3.1-nemoguard-8b-content-safety/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+nano-12b-v2-finalcheckpoint/tokenizer.json filter=lfs diff=lfs merge=lfs -text

EuroLLM-9B/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

EuroLLM-9B/README.md

@@ -0,0 +1,126 @@
+---
+license: apache-2.0
+language:
+- en
+- de
+- es
+- fr
+- it
+- pt
+- pl
+- nl
+- tr
+- sv
+- cs
+- el
+- hu
+- ro
+- fi
+- uk
+- sl
+- sk
+- da
+- lt
+- lv
+- et
+- bg
+- 'no'
+- ca
+- hr
+- ga
+- mt
+- gl
+- zh
+- ru
+- ko
+- ja
+- ar
+- hi
+library_name: transformers
+---
+
+# Model Card for EuroLLM-9B
+
+
+This is the model card for EuroLLM-9B. You can also check the instruction tuned version: [EuroLLM-9B-Instruct](https://huggingface.co/utter-project/EuroLLM-9B-Instruct).
+
+- **Developed by:** Unbabel, Instituto Superior Técnico, Instituto de Telecomunicações, University of Edinburgh, Aveni, University of Paris-Saclay, University of Amsterdam, Naver Labs, Sorbonne Université.
+- **Funded by:** European Union.
+- **Model type:** A 9B parameter multilingual transfomer LLM.
+- **Language(s) (NLP):** Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Arabic, Catalan, Chinese, Galician, Hindi, Japanese, Korean, Norwegian, Russian, Turkish, and Ukrainian. 
+- **License:** Apache License 2.0.
+
+## Model Details
+
+The EuroLLM project has the goal of creating a suite of LLMs capable of understanding and generating text in all European Union languages as well as some additional relevant languages.
+EuroLLM-9B is a 9B parameter model trained on 4 trillion tokens divided across the considered languages and several data sources: Web data, parallel data (en-xx and xx-en), and high-quality datasets.
+EuroLLM-9B-Instruct was further instruction tuned on EuroBlocks, an instruction tuning dataset with focus on general instruction-following and machine translation.
+
+
+### Model Description
+
+EuroLLM uses a standard, dense Transformer architecture:
+- We use grouped query attention (GQA) with 8 key-value heads, since it has been shown to increase speed at inference time while maintaining downstream performance.
+- We perform pre-layer normalization, since it improves the training stability, and use the RMSNorm, which is faster.
+- We use the SwiGLU activation function, since it has been shown to lead to good results on downstream tasks.
+- We use rotary positional embeddings (RoPE) in every layer, since these have been shown to lead to good performances while allowing the extension of the context length.
+
+For pre-training, we use 400 Nvidia H100 GPUs of the Marenostrum 5 supercomputer, training the model with a constant batch size of 2,800 sequences, which corresponds to approximately 12 million tokens, using the Adam optimizer, and BF16 precision.
+Here is a summary of the model hyper-parameters:
+|                                      |                      |
+|--------------------------------------|----------------------|
+| Sequence Length                      |      4,096           |
+| Number of Layers                     |         42           |
+| Embedding Size                       |           4,096      |
+| FFN Hidden Size                      |            12,288    |
+| Number of Heads                      |        32            |
+| Number of KV Heads (GQA)             |         8            |
+| Activation Function                  | SwiGLU               |
+| Position Encodings                   | RoPE (\Theta=10,000) |
+| Layer Norm                           | RMSNorm              |
+| Tied Embeddings                      | No                   |
+| Embedding Parameters                 | 0.524B               |
+| LM Head Parameters                   | 0.524B               |
+| Non-embedding Parameters             | 8.105B               |
+| Total Parameters                     | 9.154B               |
+
+## Run the model
+    
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    
+    model_id = "utter-project/EuroLLM-9B"
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    model = AutoModelForCausalLM.from_pretrained(model_id)
+    
+    text = "English: My name is EuroLLM. Portuguese:"
+    
+    inputs = tokenizer(text, return_tensors="pt")
+    outputs = model.generate(**inputs, max_new_tokens=20)
+    print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+
+
+## Results
+
+### EU Languages
+
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/63f33ecc0be81bdc5d903466/ob_1sLM8c7dxuwpv6AAHA.png)
+**Table 1:** Comparison of open-weight LLMs on multilingual benchmarks. The borda count corresponds to the average ranking of the models (see  ([Colombo et al., 2022](https://arxiv.org/abs/2202.03799))). For Arc-challenge, Hellaswag, and MMLU we are using Okapi datasets ([Lai et al., 2023](https://aclanthology.org/2023.emnlp-demo.28/)) which include 11 languages. For MMLU-Pro and MUSR we translate the English version with Tower ([Alves et al., 2024](https://arxiv.org/abs/2402.17733)) to 6 EU languages.  
+\* As there are no public versions of the pre-trained models, we evaluated them using the post-trained versions.
+
+The results in Table 1 highlight EuroLLM-9B's superior performance on multilingual tasks compared to other European-developed models (as shown by the Borda count of 1.0), as well as its strong competitiveness with non-European models, achieving results comparable to Gemma-2-9B and outperforming the rest on most benchmarks.
+
+### English
+
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/63f33ecc0be81bdc5d903466/EfilsW_p-JA13mV2ilPkm.png)
+
+**Table 2:** Comparison of open-weight LLMs on English general benchmarks.    
+\*  As there are no public versions of the pre-trained models, we evaluated them using the post-trained versions.
+
+The results in Table 2 demonstrate EuroLLM's strong performance on English tasks, surpassing most European-developed models and matching the performance of Mistral-7B (obtaining the same Borda count).
+
+
+## Bias, Risks, and Limitations
+
+EuroLLM-9B has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements).

EuroLLM-9B/config.json

@@ -0,0 +1,27 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "max_position_embeddings": 4096,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 42,
+  "num_key_value_heads": 8,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.40.1",
+  "use_cache": true,
+  "vocab_size": 128000
+}

EuroLLM-9B/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.40.1"
+}

EuroLLM-9B/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

EuroLLM-9B/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42957aec88804838805b335da1bbcce9d116024a20c6c4b56757648b9e348254
+size 2408875

EuroLLM-9B/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": null,
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/chat_template.jinja

@@ -0,0 +1,204 @@
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {{- content }}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/config.json

@@ -0,0 +1,70 @@
+{
+  "architectures": [
+    "NemotronHForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_nemotron_h.NemotronHConfig",
+    "AutoModel": "modeling_nemotron_h.NemotronHForCausalLM",
+    "AutoModelForCausalLM": "modeling_nemotron_h.NemotronHForCausalLM"
+  },
+  "bos_token_id": 1,
+  "chunk_size": 128,
+  "conv_kernel": 4,
+  "eos_token_id": 2,
+  "expand": 2,
+  "head_dim": 128,
+  "hidden_dropout": 0.0,
+  "hidden_size": 2688,
+  "hybrid_override_pattern": "MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME",
+  "initializer_range": 0.02,
+  "intermediate_size": 1856,
+  "layer_norm_epsilon": 1e-05,
+  "mamba_head_dim": 64,
+  "mamba_hidden_act": "silu",
+  "mamba_num_heads": 64,
+  "mamba_proj_bias": false,
+  "mamba_ssm_cache_dtype": "float32",
+  "max_position_embeddings": 262144,
+  "mlp_bias": false,
+  "mlp_hidden_act": "relu2",
+  "model_type": "nemotron_h",
+  "moe_intermediate_size": 1856,
+  "moe_shared_expert_intermediate_size": 3712,
+  "n_group": 1,
+  "n_groups": 8,
+  "n_routed_experts": 128,
+  "n_shared_experts": 1,
+  "norm_eps": 1e-05,
+  "norm_topk_prob": true,
+  "num_attention_heads": 32,
+  "num_experts_per_tok": 6,
+  "num_hidden_layers": 52,
+  "num_key_value_heads": 2,
+  "num_logits_to_keep": 1,
+  "pad_token_id": 0,
+  "partial_rotary_factor": 1.0,
+  "rescale_prenorm_residual": true,
+  "residual_in_fp32": false,
+  "rope_theta": 10000,
+  "routed_scaling_factor": 2.5,
+  "sliding_window": null,
+  "ssm_state_size": 128,
+  "tie_word_embeddings": false,
+  "time_step_floor": 0.0001,
+  "time_step_limit": [
+    0.0,
+    Infinity
+  ],
+  "time_step_max": 0.1,
+  "time_step_min": 0.001,
+  "topk_group": 1,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.55.4",
+  "use_bias": false,
+  "use_cache": true,
+  "use_conv_bias": true,
+  "use_mamba_kernels": true,
+  "vocab_size": 131072
+}

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/configuration_nemotron_h.py

@@ -0,0 +1,262 @@
+# coding=utf-8
+# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""NemotronH model configuration"""
+
+import re
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronHConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`NemotronHModel`]. It is used to instantiate a
+    NemotronH model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the NemotronH-v0.1 model.
+
+    [todo](todo)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 131072):
+            Vocabulary size of the NemotronH model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`NemotronHModel`]
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
+            model has a output word embedding layer.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 21504):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 52):
+            Number of hidden layers in the Transformer encoder.
+        hybrid_override_pattern (`str`, *optional*, defaults to `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
+            The pattern of the hybrid model. The pattern is a string of characters where each character represents M: Mamba2, *: Attention, -: MLP
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each attention head.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used.
+        mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
+            The non-linear activation function in the MLP layers.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in attention layers.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in MLP layers.
+        use_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the model.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        residual_in_fp32 (`bool`, *optional*, defaults to `False`):
+            Whether or not residuals should be in `float32`. If set to `False` residuals will keep the same `dtype` as the rest of the model.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
+            Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
+            integer value, only last `num_logits_to_keep` logits will be calculated.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        sliding_window (`int`, *optional*, defaults to None):
+            Sliding window attention window size.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        hidden_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the hidden states.
+        use_mamba_kernels (`bool`, *optional*, defaults to `True`):
+            Flag indicating whether or not to use the fast mamba kernels. These are available only if `mamba-ssm` and
+            `causal-conv1d` are installed, and the mamba modules are running on a CUDA device.
+        ssm_state_size (`int`, *optional*, defaults to 128):
+            The dimension of the mamba state space latents.
+        mamba_num_heads (`int`, *optional*, defaults to 128):
+            Number of heads in Mamba layers.
+        mamba_n_groups (`int`, *optional*, defaults to 8):
+            Number of groups in Mamba layers.
+        mamba_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of each Mamba head.
+        mamba_d_conv (`int`, *optional*, defaults to 4):
+            The size of the mamba convolution kernel.
+        mamba_expand (`int`, *optional*, defaults to 2):
+            Expanding factor used to determine the mamba intermediate size.
+        mamba_hidden_act (`str`, *optional*, defaults to "silu"):
+            The non-linear activation function in the Mamba layers.
+        mamba_dt_min (`float`, *optional*, defaults to 0.001):
+            Minimum value for the time step in Mamba.
+        mamba_dt_max (`float`, *optional*, defaults to 0.1):
+            Maximum value for the time step in Mamba.
+        mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
+            Limits for the time step in Mamba.
+        mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
+            Floor value for time step initialization in Mamba.
+        mamba_conv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to use bias in the convolution layer of the mamba mixer block.
+        mamba_proj_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the input and output projections of the mamba mixer block.
+        mamba_chunk_size (`int`, *optional*, defaults to 256):
+            Size of chunks for Mamba processing.
+        rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the pre-normalization residual connections.
+    """
+
+    model_type = "nemotron_h"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=131072,
+        tie_word_embeddings=False,
+        hidden_size=4096,
+        intermediate_size=21504,
+        num_hidden_layers=52,
+        hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
+        num_attention_heads=32,
+        head_dim=128,
+        num_key_value_heads=8,  # nemo: num_query_groups
+        mlp_hidden_act="relu2",
+        attention_bias=False,
+        mlp_bias=False,
+        use_bias=False,
+        initializer_range=0.02, # nemo: init_method_std
+        layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon
+        residual_in_fp32=False,  #  Megatron Core default value
+        use_cache=True,
+        num_logits_to_keep=1,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        sliding_window=None,
+        max_position_embeddings=4096,
+        attention_dropout=0.0,
+        hidden_dropout=0.0, # * ADDED
+        use_mamba_kernels=True,
+        ssm_state_size=128, # mamba_state_size
+        mamba_num_heads=128,
+        mamba_n_groups=8,  # nemo: mamba_ssm_ngroups = num_heads
+        mamba_head_dim=64,
+        mamba_d_conv=4,
+        mamba_expand=2,
+        mamba_hidden_act="silu",
+        mamba_dt_min=0.001,
+        mamba_dt_max=0.1,
+        mamba_dt_limit=(0.0, float("inf")),
+        mamba_dt_init_floor=1e-4,
+        mamba_conv_bias=True,
+        mamba_proj_bias=False,
+        mamba_chunk_size=128,
+        rescale_prenorm_residual=True,
+        n_routed_experts=8,
+        n_shared_experts=1,
+        moe_intermediate_size=7688,
+        moe_shared_expert_intermediate_size=7688,
+        num_experts_per_tok=2,
+        routed_scaling_factor=1.0,
+        n_group=1,
+        topk_group=1,
+        norm_topk_prob=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.tie_word_embeddings = tie_word_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hybrid_override_pattern = hybrid_override_pattern
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim
+        self.sliding_window = sliding_window
+        self.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.hidden_dropout = hidden_dropout
+
+        # Validate hybrid_override_pattern
+        # M: Mamba2, *: Attention, -: MLP
+        assert len(self.hybrid_override_pattern) == self.num_hidden_layers, "hybrid_override_pattern must have the same length as num_hidden_layers"
+        assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), "hybrid_override_pattern must only contain characters 'M', '*', or '-'"
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.mlp_hidden_act = mlp_hidden_act
+        self.attention_bias = attention_bias
+        self.mlp_bias = mlp_bias
+        self.use_bias = use_bias
+        self.initializer_range = initializer_range
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.residual_in_fp32 = residual_in_fp32
+
+        self.use_cache = use_cache
+        self.num_logits_to_keep = num_logits_to_keep
+
+        self.use_mamba_kernels = use_mamba_kernels
+        self.n_groups = mamba_n_groups
+        self.mamba_head_dim = mamba_head_dim
+        self.ssm_state_size = ssm_state_size
+        self.mamba_num_heads = mamba_num_heads
+        self.conv_kernel = mamba_d_conv
+        self.expand = mamba_expand
+        self.mamba_hidden_act = mamba_hidden_act
+        self.time_step_min = mamba_dt_min
+        self.time_step_max = mamba_dt_max
+        self.time_step_limit = mamba_dt_limit
+        self.time_step_floor = mamba_dt_init_floor
+        self.use_conv_bias = mamba_conv_bias
+        self.mamba_proj_bias = mamba_proj_bias
+        self.chunk_size = mamba_chunk_size
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+        self.n_routed_experts = n_routed_experts
+        self.n_shared_experts = n_shared_experts
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.routed_scaling_factor = routed_scaling_factor
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.norm_topk_prob = norm_topk_prob
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def layers_block_type(self):
+        return [
+            "mamba" if self.hybrid_override_pattern[i] == "M" else
+            "attention" if self.hybrid_override_pattern[i] == "*" else
+            "mlp" if self.hybrid_override_pattern[i] == "-" else "moe"
+            for i in range(self.num_hidden_layers)]

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "do_sample": true,
+  "bos_token_id": 1,
+  "eos_token_id": [2, 11],
+  "pad_token_id": 0,
+  "temperature": 1.0,
+  "top_p": 1.0,
+  "transformers_version": "4.55.4"
+}

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00001-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c77b0f1717f1fb11791fb62fc57ca56f59fd1427ac466849ef9705ac90729ea
+size 4991205008

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00002-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e3de804d8c8bc6607a86d486f47301822a17f274e3c54425e71ff3516cde9b6
+size 4992601472

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00003-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e113d2a3f81515599744eab31a7ea4f6cb4e6fc2089fedeb22137e14ee792c9f
+size 4992601824

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00004-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c64af357042231114495897474859e515c7d9ac00a7819ecf57d634ad8753ec5
+size 4995693256

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00005-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1aa5867c6483ac2d52891e5cdce00ee49840c2c33709f3242b05e4682b39ead0
+size 4980545984

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00006-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd411a714fad4954ee87fa76554ef79d3c85d309aff83c88b758061bf46009f1
+size 4999410040

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00007-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d16bc0bd0521e93b799e66ec913b2548417578a5f290f7023c4045dcd002f647
+size 4992601952

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00008-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc0aea38d897f28b9cc506d0fa2c2ae040562c185691d11351478841f1f474cb
+size 4992601976

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00009-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34b4715b5765fdc8fb496573a4c6c8536ad426c40d266a47d0ce1f22de441c3f
+size 4995693256

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00010-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4124abfaa922336fa8a6ba1b8f55010caac4d62a24e5990e1a819266cedcd494
+size 4992601976

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00011-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3bf1c127982233bef8ac299d5359f30aa0eaf013f8fca0645873b8e29393719
+size 4995693256

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00012-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4abbf8125860c87189dc4f37625fdba5b0c51af52eb2644f70836d9a4776f169
+size 4995693272

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/model-00013-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9458d10c7e999db805c5fa6ffa778cc0dc63478ea4210a942759358736bebf1d
+size 3239751000

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/modeling_nemotron_h.py

@@ -0,0 +1,1731 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch NemotronH model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import DynamicCache  # we need __iter__ and __len__ of pkv
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+)
+from transformers.utils.import_utils import (
+    is_causal_conv1d_available,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    is_mamba_2_ssm_available,    
+)
+from .configuration_nemotron_h import NemotronHConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.mamba.modeling_mamba2.modeling_mamba2.py with MAMBA2->NEMOTRONH,Mamba2->NemotronH
+# For Mamba2 components Mamba2->NemotronHMamba2
+if is_mamba_2_ssm_available():
+    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
+else:
+    mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined, selective_state_update = None, None, None
+
+try:
+    #from mamba_ssm.ops.triton.layernorm_gated import RMSNorm as RMSNormGated
+    from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn
+except ImportError:
+    raise ImportError("mamba-ssm is required by the Mamba model but cannot be imported")
+
+if is_causal_conv1d_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+else:
+    causal_conv1d_update, causal_conv1d_fn = None, None
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+is_fast_path_available = all(
+    (
+        selective_state_update,
+        mamba_chunk_scan_combined,
+        mamba_split_conv1d_scan_combined,
+        causal_conv1d_fn,
+        causal_conv1d_update,
+    )
+)
+
+
+_CHECKPOINT_FOR_DOC = "nvidia/Nemotron-H-56B-Base-8K"
+_CONFIG_FOR_DOC = "NemotronHConfig"
+
+
+# Helper methods for segment sum computation
+
+
+def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
+    """
+    Padding x tensor with `pad_size` on the seq_len dim (dim=1)
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)
+
+    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)
+
+
+def reshape_into_chunks(input_tensor, pad_size, chunk_size):
+    """
+    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
+    simultaneously splitting it into chunk sequences.
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
+    input_tensor = pad_tensor_by_size(input_tensor, pad_size)
+
+    if len(input_tensor.shape) == 3:
+        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
+        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
+    else:
+        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
+        return input_tensor.reshape(
+            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
+        )
+
+
+def segment_sum(input_tensor):
+    """
+    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
+    """
+    chunk_size = input_tensor.size(-1)
+    # 1. expand input tensor to have an additional dimension and repeat along that dimension
+    # [..., chunk_size] -> [..., chunk_size, chunk_size]
+    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
+    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
+    input_tensor = input_tensor.masked_fill(~mask, 0)
+    # 3. compute actual cumsum
+    tensor_segsum = torch.cumsum(input_tensor, dim=-2)
+
+    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
+    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
+    return tensor_segsum
+
+
+def apply_mask_to_padding_states(hidden_states, attention_mask):
+    """
+    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
+    """
+    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+        dtype = hidden_states.dtype
+        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+    return hidden_states
+
+# Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+class HybridMambaAttentionDynamicCache(DynamicCache):
+    """
+    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
+    (which has a constant shape regardless of seq_len).
+
+    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
+    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
+    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
+    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
+    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
+    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
+    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
+    """
+
+    def __init__(self, config, batch_size, dtype=torch.float16, device=None):
+        super().__init__()
+        self.dtype = dtype
+        self.hybrid_override_pattern = config.hybrid_override_pattern
+        self.has_previous_state = False  # only used by mamba
+        intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        ssm_state_size = config.ssm_state_size
+        conv_kernel_size = config.conv_kernel
+        self.conv_states = []
+        self.ssm_states = []
+        self.transformer_layers = []
+        for i in range(config.num_hidden_layers):
+            if self.hybrid_override_pattern[i] == "M":
+                # Mamba layer
+                self.conv_states += [
+                    torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
+                ]
+                self.ssm_states += [
+                    torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
+                ]
+            else:
+                # Attention or MLP layer
+                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.transformer_layers.append(i)
+
+        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Update the cache
+        if self.key_cache[layer_idx].shape[-1] == 0:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+        else:
+            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
+            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
+
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        for layer_idx in range(len(self.key_cache)):
+            device = self.key_cache[layer_idx].device
+            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.value_cache[layer_idx].device
+            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
+
+            device = self.conv_states[layer_idx].device
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.ssm_states[layer_idx].device
+            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
+        if len(self.key_cache) <= layer_idx:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    @classmethod
+    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    # Copied from modeling_mamba2.py
+    def update_conv_state(
+        self, layer_idx: int, new_conv_state: torch.Tensor, cache_init: bool = False
+    ) -> torch.Tensor:
+        if cache_init:
+            self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
+        else:
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
+            self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
+        return self.conv_states[layer_idx]
+
+    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
+        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
+        return self.ssm_states[layer_idx]
+
+    def reset(self):
+        self.conv_states.zero_()
+        self.ssm_states.zero_()
+
+class MambaRMSNormGated(torch.nn.Module):
+    def __init__(self, hidden_size, group_size, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+        self.group_size = group_size
+
+    # jan28b version
+    def forward(self, hidden_states, gate=None):
+        return rmsnorm_fn(x=hidden_states,
+                          weight=self.weight,
+                          bias=None, # No bias
+                          z=gate,
+                          eps=self.variance_epsilon,
+                          group_size=self.group_size,
+                          norm_before_gate=False
+        )
+
+class NemotronHMamba2Mixer(nn.Module):
+    """
+    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
+    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
+    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
+    and is why Mamba is called **selective** state spaces)
+    """
+
+    def __init__(self, config: NemotronHConfig, layer_idx: int):
+        super().__init__()
+        self.num_heads = config.mamba_num_heads
+        self.hidden_size = config.hidden_size
+        self.ssm_state_size = config.ssm_state_size
+        self.conv_kernel_size = config.conv_kernel
+        self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        self.layer_idx = layer_idx
+        self.use_conv_bias = config.use_conv_bias
+        self.activation = config.mamba_hidden_act
+        self.act = ACT2FN[config.mamba_hidden_act]
+
+        self.layer_norm_epsilon = config.layer_norm_epsilon
+
+        self.n_groups = config.n_groups
+        self.head_dim = config.mamba_head_dim
+        self.chunk_size = config.chunk_size
+
+        self.time_step_limit = config.time_step_limit
+        self.time_step_min = config.time_step_min
+        self.time_step_max = config.time_step_max
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=config.use_conv_bias,
+            kernel_size=config.conv_kernel,
+            groups=self.conv_dim,
+            padding=config.conv_kernel - 1,
+        )
+
+        # projection of the input hidden states
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size,
+            projection_size,
+            bias=config.use_bias,
+        )
+        # selective projection used to make dt, B and C input dependant
+
+        # time step projection (discretization)
+        # instantiate once and copy inv_dt in init_weights of PretrainedModel
+        self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
+
+        # S4D real initialization. These are not discretized!
+        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
+        A = torch.arange(1, self.num_heads + 1)
+        self.A_log = nn.Parameter(torch.log(A))
+        self.A_log._no_weight_decay = True
+        self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon, group_size=self.intermediate_size // self.n_groups)
+        self.D = nn.Parameter(torch.ones(self.num_heads))
+        self.D._no_weight_decay = True
+
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.use_bias = config.use_bias
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
+                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
+                " https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def cuda_kernels_forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        # 1. Gated MLP's linear projection
+        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+        projected_states = self.in_proj(hidden_states)
+
+        # Set up dimensions for reshapes later
+        batch_size, seq_len, _ = hidden_states.shape
+        groups_time_state_size = self.n_groups * self.ssm_state_size
+        d_mlp = (
+            projected_states.shape[-1]
+            - 2 * self.intermediate_size
+            - 2 * self.n_groups * self.ssm_state_size
+            - self.num_heads
+        ) // 2
+
+        # Single step calculations via cache
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
+                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+            )
+
+            # 2. Convolution sequence transformation
+            hidden_states_B_C = causal_conv1d_update(
+                hidden_states_B_C,
+                cache_params.conv_states[self.layer_idx],
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                self.activation,
+            )
+
+            hidden_states, B, C = torch.split(
+                hidden_states_B_C,
+                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                dim=-1,
+            )
+
+            # 3. SSM transformation
+            A = -torch.exp(self.A_log.float())  # (nheads,)
+            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
+            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
+            D = self.D[:, None, ...].expand(-1, self.head_dim)
+            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
+            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
+            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
+            hidden_states = selective_state_update(
+                cache_params.ssm_states[self.layer_idx],
+                hidden_states_reshaped,
+                dt,
+                A,
+                B,
+                C,
+                D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+            )
+            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
+            hidden_states = self.norm(hidden_states, gate)
+
+            # 4. Final linear projection
+            out = self.out_proj(hidden_states)[:, None, ...]
+
+        # Fused calculations or step by step if no initialized cache is found
+        else:
+            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
+            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
+
+            # 2-4. Fused kernel for conv1d, SSM, and the final projection
+            if self.training and cache_params is None:
+                out = mamba_split_conv1d_scan_combined(
+                    projected_states,
+                    self.conv1d.weight.squeeze(1),
+                    self.conv1d.bias,
+                    self.dt_bias,
+                    A,
+                    D=self.D,
+                    chunk_size=self.chunk_size,
+                    seq_idx=None,  # was seq_idx
+                    activation=self.activation,
+                    rmsnorm_weight=self.norm.weight,
+                    rmsnorm_eps=self.norm.variance_epsilon,
+                    outproj_weight=self.out_proj.weight,
+                    outproj_bias=self.out_proj.bias,
+                    headdim=self.head_dim,
+                    ngroups=self.n_groups,
+                    norm_before_gate=False,
+                    return_final_states=False,
+                    **dt_limit_kwargs,
+                )
+
+            else:
+                _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+                )
+
+                # 2. Convolution sequence transformation
+                # Init cache
+                if cache_params is not None:
+                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                    conv_states = nn.functional.pad(
+                        hidden_states_B_C_transposed,
+                        (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
+                    )
+                    cache_params.update_conv_state(
+                        layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
+                    )
+
+                if self.activation not in ["silu", "swish"]:
+                    hidden_states_B_C = self.act(
+                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
+                    )
+                else:
+                    hidden_states_B_C = causal_conv1d_fn(
+                        x=hidden_states_B_C.transpose(1, 2),
+                        weight=self.conv1d.weight.squeeze(1),
+                        bias=self.conv1d.bias,
+                        activation=self.activation,
+                    ).transpose(1, 2)
+                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+                hidden_states, B, C = torch.split(
+                    hidden_states_B_C,
+                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                    dim=-1,
+                )
+
+                # 3. SSM transformation
+                scan_output, ssm_state = mamba_chunk_scan_combined(
+                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
+                    dt,
+                    A,
+                    B.view(batch_size, seq_len, self.n_groups, -1),
+                    C.view(batch_size, seq_len, self.n_groups, -1),
+                    chunk_size=self.chunk_size,
+                    D=self.D,
+                    z=None,
+                    seq_idx=None,
+                    return_final_states=True,
+                    dt_bias=self.dt_bias,
+                    dt_softplus=True,
+                    **dt_limit_kwargs,
+                )
+
+                # Init cache
+                if ssm_state is not None and cache_params is not None:
+                    cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+                scan_output = scan_output.view(batch_size, seq_len, -1)
+
+                # Multiply "gate" branch and apply extra normalization layer
+                scan_output = self.norm(scan_output, gate)
+
+                # 4. Final linear projection
+                out = self.out_proj(scan_output)
+        return out
+
+    # fmt: off
+    def torch_forward(self, input_states, cache_params: Optional[HybridMambaAttentionDynamicCache]=None, cache_position:Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
+
+        # 1. Gated MLP's linear projection
+        input_states = apply_mask_to_padding_states(input_states, attention_mask)
+        projected_states = self.in_proj(input_states)
+        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
+        _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
+        )
+
+        # 2. Convolution sequence transformation
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
+
+            # We need to guarantee that anything regarding the cache is on the same device
+            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
+
+            hidden_states_B_C = torch.sum(
+                conv_states * self.conv1d.weight.squeeze(1), dim=-1
+            )
+            if self.use_conv_bias:
+                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
+            hidden_states_B_C = self.act(hidden_states_B_C)
+        else:
+            # Init cache
+            if cache_params is not None:
+                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                conv_states = nn.functional.pad(
+                    hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
+                )
+                cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
+
+            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))
+
+        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+        hidden_states, B, C = torch.split(
+            hidden_states_B_C,
+            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
+            dim=-1
+        )
+
+        # 3. SSM transformation
+        A = -torch.exp(self.A_log.float())                            # [num_heads]
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            # We need to guarantee that anything regarding the cache is on the same device
+            cache_device = cache_params.ssm_states.device
+
+            # Note: there is no need to pad parameter matrices here, as there is just one new token
+            # for batched generation
+            dt = dt[:, 0, :][:, None, ...]
+            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
+            # [num_heads] -> [num_heads, head_dim]
+            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
+
+            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            # [bsz, num_heads, head_dim, state_size]
+            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
+
+            # Discretize B
+            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
+            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
+            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
+            B = B.reshape(batch_size, -1, B.shape[-1])
+            # [bsz, num_heads, head_dim, state_size]
+            dB = dt[..., None] * B[..., None, :]
+
+            # Discretize x into dB
+            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
+            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
+            dBx = (dB * hidden_states[..., None]).to(device=cache_device)
+
+            # State calculation
+            cache_params.update_ssm_state(
+                layer_idx=self.layer_idx,
+                new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
+            )
+
+            # Subsequent output
+            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
+            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
+            C = C.reshape(batch_size, -1, C.shape[-1])
+            # [bsz, num_heads, head_dim]
+
+            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
+            # Reshape ssm_states to merge the first two dimensions
+            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
+            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
+            y = torch.bmm(ssm_states_reshaped, C_reshaped)
+            y = y.view(batch_size, self.num_heads, self.head_dim)
+
+            # D skip connection
+            # [num_heads] -> [num_heads, head_dim]
+            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
+            y = (y + hidden_states * D).to(y.dtype)
+
+            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
+            y = y.reshape(batch_size, -1)[:, None, ...]
+        else:
+            # begin ssd naive implementation without einsums
+            dt = nn.functional.softplus(dt + self.dt_bias)
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
+            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            B = B.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            C = C.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
+
+            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
+
+            # Discretize x and A
+            hidden_states = hidden_states * dt[..., None]
+            A = A.to(hidden_states.dtype) * dt
+
+            # Rearrange into blocks/chunks
+            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
+
+            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
+            A = A.permute(0, 3, 1, 2)
+            A_cumsum = torch.cumsum(A, dim=-1)
+
+            # 1. Compute the output for each intra-chunk (diagonal blocks)
+            # This is the analog of a causal mask
+            L = torch.exp(segment_sum(A))
+
+            # Contraction of C and B to get G (attention-weights like)
+            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]  # shape: (b, c, l, s, h, n)
+            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)
+
+            # Compute M, equivalent to applying attention mask to weights
+            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
+            M = M_intermediate.sum(dim=-1)
+
+            # Compute Y_diag (apply to values)
+            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
+
+            # 2. Compute the state for each intra-chunk
+            # (right term of low-rank factorization of off-diagonal blocks; B terms)
+            decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
+            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
+            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
+
+            # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
+            # (middle term of factorization of off-diag blocks; A terms)
+            if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
+            else:
+                previous_states = torch.zeros_like(states[:, :1])
+            states = torch.cat([previous_states, states], dim=1)
+            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
+            states, ssm_state = new_states[:, :-1], new_states[:, -1]
+
+            # 4. Compute state -> output conversion per chunk
+            # (left term of low-rank factorization of off-diagonal blocks; C terms)
+            state_decay_out = torch.exp(A_cumsum)
+            C_times_states = (C[..., None, :] * states[:, :, None, ...])
+            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
+            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
+
+            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
+            y = Y_diag + Y_off
+            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
+            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
+
+            y = y + D_residual
+            # Cutting off padded chunks
+            if pad_size > 0:
+                y = y[:, :seq_len, :, :]
+            y = y.reshape(batch_size, seq_len, -1)
+
+            # Init cache
+            if ssm_state is not None and cache_params is not None:
+                cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+        scan_output = self.norm(y, gate)
+
+        # end ssd naive
+
+        # 4. Final linear projection
+        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
+        return contextualized_states
+    # fmt: on
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
+            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
+        dtype = hidden_states.dtype
+        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
+            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)
+
+
+class NemotronHRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        NemotronHRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        # Weights are in float32
+        return (self.weight.to(torch.float32) * hidden_states).to(input_dtype)
+
+class NemotronHBlock(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.residual_in_fp32 = config.residual_in_fp32
+        self.norm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        # M: Mamba2, *: Attention, -: MLP
+        self.block_type = config.layers_block_type[layer_idx]
+        if self.block_type == "mamba":
+            self.mixer = NemotronHMamba2Mixer(config, layer_idx=layer_idx)
+        elif self.block_type == "attention":
+            self.mixer = NEMOTRONH_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+        elif self.block_type == "mlp":
+            self.mixer = NemotronHMLP(config, layer_idx=layer_idx)
+        elif self.block_type == "moe":
+            self.mixer = NemotronHMOE(config, layer_idx=layer_idx)
+        else:
+            raise ValueError(f"Invalid layer pattern {config.hybrid_override_pattern[layer_idx]}")
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        with torch.cuda.stream(torch.cuda.default_stream(hidden_states.device)):
+            # * Use torch.cuda.stream() to avoid NaN issues when using multiple GPUs
+            residual = hidden_states
+            hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
+            if self.residual_in_fp32:
+                residual = residual.to(torch.float32)
+
+            if self.block_type == "mamba":
+                hidden_states = self.mixer(
+                    hidden_states, cache_params=cache_params, cache_position=cache_position
+                )
+            elif self.block_type == "attention":
+                hidden_states = self.mixer(
+                    hidden_states, cache_position=cache_position
+                )
+                hidden_states = hidden_states[0]
+            elif self.block_type in ["mlp", "moe"]:
+                hidden_states = self.mixer(
+                    hidden_states
+                )
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            hidden_states = residual + hidden_states
+            return hidden_states
+
+
+# Copied from transformers.models.nemotron.modeling_nemotron Nemotron->NemotronH
+class NemotronHMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.intermediate_size
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.mlp_hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.up_proj(x)))
+
+
+class NemotronHMOE(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.experts = nn.ModuleList(
+            [
+                NemotronHMLP(config, intermediate_size=config.moe_intermediate_size, layer_idx=layer_idx)
+                for _ in range(config.n_routed_experts)
+            ]
+        )
+        self.gate = NemotronHTopkRouter(config)
+        self.shared_experts = NemotronHMLP(
+            config=config, intermediate_size=config.moe_shared_expert_intermediate_size, layer_idx=layer_idx
+        )
+
+    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
+        r"""
+        CALL FOR CONTRIBUTION! I don't have time to optimise this right now, but expert weights need to be fused
+        to not have to do a loop here (deepseek has 256 experts soooo yeah).
+        """
+        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)
+        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts))
+        expert_mask = expert_mask.permute(2, 0, 1)
+
+        for expert_idx in range(len(self.experts)):
+            expert = self.experts[expert_idx]
+            mask = expert_mask[expert_idx]
+            token_indices, weight_indices = torch.where(mask)
+
+            if token_indices.numel() > 0:
+                expert_weights = topk_weights[token_indices, weight_indices]
+                expert_input = hidden_states[token_indices]
+                expert_output = expert(expert_input)
+                weighted_output = expert_output * expert_weights.unsqueeze(-1)
+                final_hidden_states.index_add_(0, token_indices, weighted_output)
+
+        # in original deepseek, the output of the experts are gathered once we leave this module
+        # thus the moe module is itelsf an IsolatedParallel module
+        # and all expert are "local" meaning we shard but we don't gather
+        return final_hidden_states.type(hidden_states.dtype)
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        topk_indices, topk_weights = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class NemotronHTopkRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size), dtype=torch.float32))
+        self.register_buffer("e_score_correction_bias", torch.zeros(self.n_routed_experts, dtype=torch.float32))
+
+    @torch.no_grad()
+    def get_topk_indices(self, scores):
+        scores_for_choice = scores.view(-1, self.n_routed_experts) + self.e_score_correction_bias.unsqueeze(0)
+        group_scores = (
+            scores_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        return topk_indices
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        scores = router_logits.sigmoid()
+        topk_indices = self.get_topk_indices(scores)
+        topk_weights = scores.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class NemotronHAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: NemotronHConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        if hasattr(config, "head_dim") and config.head_dim is not None:
+            self.head_dim = config.head_dim
+        else:
+            self.head_dim = config.hidden_size // self.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.head_dim * self.num_heads, self.hidden_size, bias=config.attention_bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        # position_embeddings: Tuple[torch.Tensor, torch.Tensor], #TODO
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        #attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.num_heads * self.head_dim)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
+#class JambaFlashAttention2(JambaAttention):
+class NemotronHFlashAttention2(NemotronHAttention):
+    """
+    Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        #attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
+#class JambaSdpaAttention(JambaAttention):
+class NemotronHSdpaAttention(NemotronHAttention):
+    """
+    Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from NemotronHAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "NemotronHModel is using NemotronHSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+        is_causal = True if self.is_causal and causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+NEMOTRONH_ATTENTION_CLASSES = {
+    "eager": NemotronHAttention,
+    "flash_attention_2": NemotronHFlashAttention2,
+    "sdpa": NemotronHSdpaAttention,
+}
+
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2PreTrainedModel
+class NemotronHPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = NemotronHConfig
+    base_model_prefix = "backbone"
+    _no_split_modules = ["NemotronHBlock"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, NemotronHMamba2Mixer):
+            module.A_log._no_weight_decay = True
+            module.D._no_weight_decay = True
+
+            dt = torch.exp(
+                torch.rand(self.config.mamba_num_heads)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+
+            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                module.dt_bias.copy_(inv_dt)
+            module.dt_bias._no_reinit = True
+
+        if isinstance(module, nn.Linear):
+            if module.bias is not None:
+                if not getattr(module.bias, "_no_reinit", False):
+                    nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=self.config.initializer_range)
+
+        # TODO: Check
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            for name, p in module.named_parameters():
+                if name in ["out_proj.weight"]:
+                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                    # We need to reinit p since this code could be called multiple times
+                    # Having just p *= scale would repeatedly scale it down
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(self.config.num_hidden_layers)
+
+
+@dataclass
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2Output with MAMBA2->NemotronH,Mamba2->NemotronH
+class NemotronHOutput(ModelOutput):
+    """
+    Class for the NemotronH model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.mamba2.modeling_mamba2.MambaCausalLMOutput with Mamba2->NemotronH
+class NemotronHCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+NEMOTRONH_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`NemotronHConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+NEMOTRONH_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary.
+
+            If `cache_params.seqlen_offset>0`, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        position_ids (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        cache_params (`HybridMambaAttentionDynamicCache`, *optional*):
+            If passed along, the model uses the previous state in all the blocks (which will give the output for the
+            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
+        use_cache (`bool`, *optional*):
+            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            The position of the current input in the cache. This is used to ensure that the cache is correctly updated.
+            If `cache_params` is passed, `cache_position` should also be passed.
+        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+"""
+
+
+@add_start_docstrings(
+    "The bare NemotronH Model transformer outputting raw hidden-states without any specific head on top.",
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHModel(NemotronHPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([NemotronHBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        # Initialize weights and apply final processing
+        self._register_load_state_dict_pre_hook(self.load_hook)
+        self.post_init()
+
+    def load_hook(self, state_dict, prefix, *args):
+        for k in state_dict:
+            if "embedding." in k:
+                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
+                break
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, NemotronHOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        # use_cache = use_cache if use_cache is not None else self.config.use_cache
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # From zamba_modeling.py
+        if use_cache and cache_params is None:
+            logger.warning_once(
+                "NemotronH requires an initialized `NemotronHHybridDynamicCache` to return a cache. None was "
+                "provided, so no cache will be returned."
+            )
+
+        hidden_states = inputs_embeds
+
+        if cache_position is None:
+            cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
+        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        # Until HERE
+
+        for layer_idx, mixer_block in enumerate(self.layers):
+            # Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
+            if mixer_block.block_type == "mamba":
+                layer_mask = mamba_mask
+            elif mixer_block.block_type == "attention":
+                layer_mask = causal_mask
+            elif mixer_block.block_type in ["mlp", "moe"]:
+                layer_mask = None
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__, hidden_states, cache_params, cache_position, layer_mask
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    cache_position=cache_position,
+                    attention_mask=layer_mask,
+                )
+
+            # TODO: Store attentions
+            # if output_attentions:
+            #     if layer_outputs[1] is not None:
+            #         # append attentions only of attention layers. Mamba layers return `None` as the attention weights
+            #         all_self_attns += (layer_outputs[1],)
+
+            # TODO (Check): should it happen before the forward pass?
+            # if output_hidden_states:
+            #     all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return NemotronHOutput(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    # Copied from transformers.models.jamba.modeling_jamba.JambaModel._update_causal_mask
+    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        target_length = cache_position[-1] + 1
+
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    def _update_mamba_mask(self, attention_mask, cache_position):
+        """
+        No need for zeroing states when
+            1. Cached forward
+            2. Attending to all inputs
+        """
+        mamba_mask = attention_mask
+        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
+            mamba_mask = None
+        return mamba_mask
+
+
+@add_start_docstrings(
+    """
+    The NEMOTRONH Model transformer with a language modeling head on top (linear layer with weights not tied to the input
+    embeddings).
+    """,
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHForCausalLM(NemotronHPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = NemotronHModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+        # Overwitten -- uses `cache_params` as opposed to `past_key_values`
+        empty_past_kv = past_key_values is None
+
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
+        #              (we can't check exception 3 while compiling)
+        if not empty_past_kv:
+            if (
+                inputs_embeds is not None  # Exception 1
+                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
+            ):
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+        else:
+            past_key_values = HybridMambaAttentionDynamicCache(
+                self.config, input_ids.shape[0], self.dtype, device=self.device
+            )
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if not empty_past_kv:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and empty_past_kv:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+                "logits_to_keep": self.config.num_logits_to_keep,
+                "cache_position": cache_position,
+            }
+        )
+        return model_inputs
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHCausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,  # for now we need this for generation
+    ) -> Union[Tuple, NemotronHCausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        nemotron_h_outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = nemotron_h_outputs[0]
+
+        # TODO: Check zamba_modeling.py: https://github.com/huggingface/transformers/blob/d7188ba600e36d3fd191b12e19f1b3bb81a8404f/src/transformers/models/zamba/modeling_zamba.py#L1284C1-L1286C2
+        #logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+        logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + nemotron_h_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return NemotronHCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=nemotron_h_outputs.cache_params,
+            hidden_states=nemotron_h_outputs.hidden_states,
+            attentions=nemotron_h_outputs.attentions,
+        )

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/nano_v3_reasoning_parser.py

@@ -0,0 +1,19 @@
+from vllm.reasoning.abs_reasoning_parsers import ReasoningParserManager
+from vllm.reasoning.deepseek_r1_reasoning_parser import DeepSeekR1ReasoningParser
+
+
+@ReasoningParserManager.register_module("nano_v3")
+class NanoV3ReasoningParser(DeepSeekR1ReasoningParser):
+    def extract_reasoning(self, model_output, request):
+        reasoning_content, final_content = super().extract_reasoning(
+            model_output, request
+        )
+        if (
+            hasattr(request, "chat_template_kwargs")
+            and request.chat_template_kwargs
+            and request.chat_template_kwargs.get("enable_thinking") is False
+            and final_content is None
+        ):
+            reasoning_content, final_content = final_content, reasoning_content
+
+        return reasoning_content, final_content

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-new-chat-template/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6021eb6847e682f89aa52d5eb6e8c7d902a23acfc8137e25211cf84828f1592
+size 17077485

NVIDIA-Nemotron-Nano-12B-v2/.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

NVIDIA-Nemotron-Nano-12B-v2/config.json

@@ -0,0 +1,57 @@
+{
+  "architectures": [
+    "NemotronHForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "head_dim": 128,
+  "auto_map": {
+    "AutoConfig": "configuration_nemotron_h.NemotronHConfig",
+    "AutoModelForCausalLM": "modeling_nemotron_h.NemotronHForCausalLM"
+  },
+  "bos_token_id": 1,
+  "chunk_size": 128,
+  "conv_kernel": 4,
+  "eos_token_id": 12,
+  "hidden_dropout": 0.0,
+  "hidden_size": 5120,
+  "hybrid_override_pattern": "M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M-",
+  "initializer_range": 0.02,
+  "intermediate_size": 20480,
+  "layer_norm_epsilon": 1e-05,
+  "mamba_head_dim": 80,
+  "mamba_hidden_act": "silu",
+  "mamba_num_heads": 128,
+  "mamba_proj_bias": false,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "mlp_hidden_act": "relu2",
+  "model_type": "nemotron_h",
+  "n_groups": 8,
+  "num_attention_heads": 40,
+  "num_hidden_layers": 62,
+  "num_key_value_heads": 8,
+  "num_logits_to_keep": 1,
+  "pad_token_id": 0,
+  "rescale_prenorm_residual": true,
+  "residual_in_fp32": false,
+  "rms_norm_eps": 1e-05,
+  "sliding_window": null,
+  "ssm_state_size": 128,
+  "tie_word_embeddings": false,
+  "time_step_floor": 0.0001,
+  "time_step_limit": [
+    0.0,
+    Infinity
+  ],
+  "time_step_max": 0.1,
+  "time_step_min": 0.001,
+  "time_step_rank": 256,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_bias": false,
+  "use_cache": true,
+  "use_conv_bias": true,
+  "use_mamba_kernels": true,
+  "vocab_size": 131072
+}

NVIDIA-Nemotron-Nano-12B-v2/configuration_nemotron_h.py

@@ -0,0 +1,245 @@
+# coding=utf-8
+# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""NemotronH model configuration"""
+
+import re
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronHConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`NemotronHModel`]. It is used to instantiate a
+    NemotronH model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the NemotronH-v0.1 model.
+
+    [todo](todo)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 131072):
+            Vocabulary size of the NemotronH model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`NemotronHModel`]
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
+            model has a output word embedding layer.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 21504):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 52):
+            Number of hidden layers in the Transformer encoder.
+        hybrid_override_pattern (`str`, *optional*, defaults to `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
+            The pattern of the hybrid model. The pattern is a string of characters where each character represents M: Mamba2, *: Attention, -: MLP
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        attention_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each attention head.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used.
+        mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
+            The non-linear activation function in the MLP layers.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in attention layers.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in MLP layers.
+        use_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the model.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        residual_in_fp32 (`bool`, *optional*, defaults to `False`):
+            Whether or not residuals should be in `float32`. If set to `False` residuals will keep the same `dtype` as the rest of the model.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
+            Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
+            integer value, only last `num_logits_to_keep` logits will be calculated.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        sliding_window (`int`, *optional*, defaults to None):
+            Sliding window attention window size.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        hidden_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the hidden states.
+        use_mamba_kernels (`bool`, *optional*, defaults to `True`):
+            Flag indicating whether or not to use the fast mamba kernels. These are available only if `mamba-ssm` and
+            `causal-conv1d` are installed, and the mamba modules are running on a CUDA device.
+        ssm_state_size (`int`, *optional*, defaults to 128):
+            The dimension of the mamba state space latents.
+        mamba_num_heads (`int`, *optional*, defaults to 128):
+            Number of heads in Mamba layers.
+        mamba_n_groups (`int`, *optional*, defaults to 8):
+            Number of groups in Mamba layers.
+        mamba_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of each Mamba head.
+        mamba_d_conv (`int`, *optional*, defaults to 4):
+            The size of the mamba convolution kernel.
+        mamba_expand (`int`, *optional*, defaults to 2):
+            Expanding factor used to determine the mamba intermediate size.
+        mamba_hidden_act (`str`, *optional*, defaults to "silu"):
+            The non-linear activation function in the Mamba layers.
+        mamba_dt_min (`float`, *optional*, defaults to 0.001):
+            Minimum value for the time step in Mamba.
+        mamba_dt_max (`float`, *optional*, defaults to 0.1):
+            Maximum value for the time step in Mamba.
+        mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
+            Limits for the time step in Mamba.
+        mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
+            Floor value for time step initialization in Mamba.
+        mamba_conv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to use bias in the convolution layer of the mamba mixer block.
+        mamba_proj_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the input and output projections of the mamba mixer block.
+        mamba_chunk_size (`int`, *optional*, defaults to 256):
+            Size of chunks for Mamba processing.
+        rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the pre-normalization residual connections.
+    """
+
+    model_type = "nemotron_h"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=131072,
+        tie_word_embeddings=False,
+        hidden_size=4096,
+        intermediate_size=21504,
+        num_hidden_layers=52,
+        hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
+        num_attention_heads=32,
+        #attention_head_dim=128,
+        head_dim=128,
+        num_key_value_heads=8,  # nemo: num_query_groups
+        mlp_hidden_act="relu2",
+        attention_bias=False,
+        mlp_bias=False,
+        use_bias=False,
+        initializer_range=0.02, # nemo: init_method_std
+        layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon
+        residual_in_fp32=False,  #  Megatron Core default value
+        use_cache=True,
+        num_logits_to_keep=1,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        sliding_window=None,
+        max_position_embeddings=4096,
+        attention_dropout=0.0,
+        hidden_dropout=0.0, # * ADDED
+        use_mamba_kernels=True,
+        ssm_state_size=128, # mamba_state_size
+        mamba_num_heads=128,
+        mamba_n_groups=8,  # nemo: mamba_ssm_ngroups = num_heads
+        mamba_head_dim=64,
+        mamba_d_conv=4,
+        mamba_expand=2,
+        mamba_hidden_act="silu",
+        mamba_dt_min=0.001,
+        mamba_dt_max=0.1,
+        mamba_dt_limit=(0.0, float("inf")),
+        mamba_dt_init_floor=1e-4,
+        mamba_conv_bias=True,
+        mamba_proj_bias=False,
+        mamba_chunk_size=256,
+        rescale_prenorm_residual=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.tie_word_embeddings = tie_word_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hybrid_override_pattern = hybrid_override_pattern
+        self.num_attention_heads = num_attention_heads
+        #self.attention_head_dim = attention_head_dim
+        self.head_dim = head_dim
+        self.sliding_window = sliding_window
+        self.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.hidden_dropout = hidden_dropout
+
+        # Validate hybrid_override_pattern
+        # M: Mamba2, *: Attention, -: MLP
+        assert len(self.hybrid_override_pattern) == self.num_hidden_layers, "hybrid_override_pattern must have the same length as num_hidden_layers"
+        assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), "hybrid_override_pattern must only contain characters 'M', '*', or '-'"
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.mlp_hidden_act = mlp_hidden_act
+        self.attention_bias = attention_bias
+        self.mlp_bias = mlp_bias
+        self.use_bias = use_bias
+        self.initializer_range = initializer_range
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.residual_in_fp32 = residual_in_fp32
+
+        self.use_cache = use_cache
+        self.num_logits_to_keep = num_logits_to_keep
+
+        self.use_mamba_kernels = use_mamba_kernels
+        self.n_groups = mamba_n_groups
+        self.mamba_head_dim = mamba_head_dim
+        self.ssm_state_size = ssm_state_size
+        self.mamba_num_heads = mamba_num_heads
+        self.conv_kernel = mamba_d_conv
+        self.expand = mamba_expand
+        self.mamba_hidden_act = mamba_hidden_act
+        self.time_step_min = mamba_dt_min
+        self.time_step_max = mamba_dt_max
+        self.time_step_limit = mamba_dt_limit
+        self.time_step_floor = mamba_dt_init_floor
+        self.use_conv_bias = mamba_conv_bias
+        self.mamba_proj_bias = mamba_proj_bias
+        self.chunk_size = mamba_chunk_size
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def layers_block_type(self):
+        return [
+            "mamba" if self.hybrid_override_pattern[i] == "M" else
+            "attention" if self.hybrid_override_pattern[i] == "*" else "mlp"
+            for i in range(self.num_hidden_layers)]

NVIDIA-Nemotron-Nano-12B-v2/generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    2,
+    11,
+    12
+  ],
+  "pad_token_id": 0,
+  "transformers_version": "4.51.3"
+}

NVIDIA-Nemotron-Nano-12B-v2/model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a5b53fe19bb584eaa8051c23e191170e29f3ecb31755be195d3d373e193f8db
+size 4830843560

NVIDIA-Nemotron-Nano-12B-v2/model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29cf35931508d965f7a4526cb1a8f5cbb4cc9d881afea8b331bf60dc8ee1cb1a
+size 4874263544

NVIDIA-Nemotron-Nano-12B-v2/model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0093760f93ae3d2c12313c17ef8efd01784c96ceebc582330354b5c59bc2941a
+size 4790388152

NVIDIA-Nemotron-Nano-12B-v2/model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a49e6ee851799b4872d6aeeb33d72d2a18a030e8dcf135c37802be555bb3ad63
+size 4874398064

NVIDIA-Nemotron-Nano-12B-v2/model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bcdc6c89f856c92af2f9c2aa474e35a4cb42a9676cc9626a049327ca41367e9
+size 3907972648

NVIDIA-Nemotron-Nano-12B-v2/model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f07483398386c8d55bac2850ea48a099427bbe72ae905fa37c5d4a3b1b598f43
+size 1342177408

NVIDIA-Nemotron-Nano-12B-v2/model.safetensors.index.json

@@ -0,0 +1,376 @@
+{
+  "metadata": {
+    "total_size": 24620002304
+  },
+  "weight_map": {
+    "backbone.embeddings.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.A_log": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.D": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.conv1d.bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.conv1d.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.dt_bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.in_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.mixer.out_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.0.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.1.mixer.down_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.1.mixer.up_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.1.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.10.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.10.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.10.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.11.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.11.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.12.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.12.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.12.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.13.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.14.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.14.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.14.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.15.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.16.mixer.k_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.16.mixer.o_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.16.mixer.q_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.16.mixer.v_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.16.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.17.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.17.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.17.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.18.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.19.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.19.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.19.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.2.mixer.A_log": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.D": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.conv1d.bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.conv1d.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.dt_bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.in_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.mixer.out_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.2.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.20.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.20.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.21.mixer.down_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.21.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.21.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.A_log": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.D": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.conv1d.bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.conv1d.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.dt_bias": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.in_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.mixer.out_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.22.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.23.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.23.mixer.up_proj.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.23.norm.weight": "model-00002-of-00006.safetensors",
+    "backbone.layers.24.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.24.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.25.mixer.k_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.25.mixer.o_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.25.mixer.q_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.25.mixer.v_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.25.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.26.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.26.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.26.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.27.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.28.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.28.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.28.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.29.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.3.mixer.down_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.3.mixer.up_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.3.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.30.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.30.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.30.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.31.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.32.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.32.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.32.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.33.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.34.mixer.k_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.34.mixer.o_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.34.mixer.q_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.34.mixer.v_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.34.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.35.mixer.down_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.35.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.35.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.A_log": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.D": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.conv1d.bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.conv1d.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.dt_bias": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.in_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.mixer.out_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.36.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.37.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.37.mixer.up_proj.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.37.norm.weight": "model-00003-of-00006.safetensors",
+    "backbone.layers.38.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.38.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.39.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.39.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.39.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.4.mixer.A_log": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.D": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.conv1d.bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.conv1d.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.dt_bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.in_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.mixer.out_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.4.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.40.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.40.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.41.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.41.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.41.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.42.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.43.mixer.k_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.43.mixer.o_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.43.mixer.q_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.43.mixer.v_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.43.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.44.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.44.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.44.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.45.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.46.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.46.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.46.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.47.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.48.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.48.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.48.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.in_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.mixer.out_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.49.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.5.mixer.down_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.5.mixer.up_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.5.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.50.mixer.down_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.50.mixer.up_proj.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.50.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.A_log": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.D": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.conv1d.bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.conv1d.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.dt_bias": "model-00004-of-00006.safetensors",
+    "backbone.layers.51.mixer.in_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.51.mixer.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.51.mixer.out_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.51.norm.weight": "model-00004-of-00006.safetensors",
+    "backbone.layers.52.mixer.k_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.52.mixer.o_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.52.mixer.q_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.52.mixer.v_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.52.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.53.mixer.down_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.53.mixer.up_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.53.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.A_log": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.D": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.conv1d.bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.conv1d.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.dt_bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.in_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.mixer.out_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.54.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.55.mixer.down_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.55.mixer.up_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.55.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.A_log": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.D": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.conv1d.bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.conv1d.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.dt_bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.in_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.mixer.out_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.56.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.57.mixer.down_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.57.mixer.up_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.57.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.A_log": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.D": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.conv1d.bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.conv1d.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.dt_bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.in_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.mixer.out_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.58.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.59.mixer.down_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.59.mixer.up_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.59.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.6.mixer.A_log": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.D": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.conv1d.bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.conv1d.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.dt_bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.in_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.mixer.out_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.6.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.60.mixer.A_log": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.D": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.conv1d.bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.conv1d.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.dt_bias": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.in_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.mixer.out_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.60.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.61.mixer.down_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.61.mixer.up_proj.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.61.norm.weight": "model-00005-of-00006.safetensors",
+    "backbone.layers.7.mixer.k_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.7.mixer.o_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.7.mixer.q_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.7.mixer.v_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.7.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.8.mixer.down_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.8.mixer.up_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.8.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.A_log": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.D": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.conv1d.bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.conv1d.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.dt_bias": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.in_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.mixer.out_proj.weight": "model-00001-of-00006.safetensors",
+    "backbone.layers.9.norm.weight": "model-00001-of-00006.safetensors",
+    "backbone.norm_f.weight": "model-00005-of-00006.safetensors",
+    "lm_head.weight": "model-00006-of-00006.safetensors"
+  }
+}

NVIDIA-Nemotron-Nano-12B-v2/modeling_nemotron_h.py

@@ -0,0 +1,1638 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch NemotronH model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import DynamicCache  # we need __iter__ and __len__ of pkv
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+)
+from transformers.utils.import_utils import (
+    is_causal_conv1d_available,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    is_mamba_2_ssm_available,
+)
+from .configuration_nemotron_h import NemotronHConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.mamba.modeling_mamba2.modeling_mamba2.py with MAMBA2->NEMOTRONH,Mamba2->NemotronH
+# For Mamba2 components Mamba2->NemotronHMamba2
+if is_mamba_2_ssm_available():
+    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
+else:
+    mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined, selective_state_update = None, None, None
+
+try:
+    #from mamba_ssm.ops.triton.layernorm_gated import RMSNorm as RMSNormGated
+    from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn
+except ImportError:
+    raise ImportError("mamba-ssm is required by the Mamba model but cannot be imported")
+
+if is_causal_conv1d_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+else:
+    causal_conv1d_update, causal_conv1d_fn = None, None
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+is_fast_path_available = all(
+    (
+        selective_state_update,
+        mamba_chunk_scan_combined,
+        mamba_split_conv1d_scan_combined,
+        causal_conv1d_fn,
+        causal_conv1d_update,
+    )
+)
+
+
+_CHECKPOINT_FOR_DOC = "nvidia/Nemotron-H-56B-Base-8K"
+_CONFIG_FOR_DOC = "NemotronHConfig"
+
+
+# Helper methods for segment sum computation
+
+
+def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
+    """
+    Padding x tensor with `pad_size` on the seq_len dim (dim=1)
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)
+
+    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)
+
+
+def reshape_into_chunks(input_tensor, pad_size, chunk_size):
+    """
+    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
+    simultaneously splitting it into chunk sequences.
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
+    input_tensor = pad_tensor_by_size(input_tensor, pad_size)
+
+    if len(input_tensor.shape) == 3:
+        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
+        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
+    else:
+        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
+        return input_tensor.reshape(
+            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
+        )
+
+
+def segment_sum(input_tensor):
+    """
+    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
+    """
+    chunk_size = input_tensor.size(-1)
+    # 1. expand input tensor to have an additional dimension and repeat along that dimension
+    # [..., chunk_size] -> [..., chunk_size, chunk_size]
+    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
+    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
+    input_tensor = input_tensor.masked_fill(~mask, 0)
+    # 3. compute actual cumsum
+    tensor_segsum = torch.cumsum(input_tensor, dim=-2)
+
+    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
+    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
+    return tensor_segsum
+
+
+def apply_mask_to_padding_states(hidden_states, attention_mask):
+    """
+    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
+    """
+    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+        dtype = hidden_states.dtype
+        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+    return hidden_states
+
+# Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+class HybridMambaAttentionDynamicCache(DynamicCache):
+    """
+    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
+    (which has a constant shape regardless of seq_len).
+
+    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
+    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
+    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
+    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
+    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
+    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
+    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
+    """
+
+    def __init__(self, config, batch_size, dtype=torch.float16, device=None):
+        super().__init__()
+        self.dtype = dtype
+        self.hybrid_override_pattern = config.hybrid_override_pattern
+        self.has_previous_state = False  # only used by mamba
+        #intermediate_size = config.expand * config.hidden_size
+        intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        ssm_state_size = config.ssm_state_size
+        conv_kernel_size = config.conv_kernel
+        self.conv_states = []
+        self.ssm_states = []
+        self.transformer_layers = []
+        for i in range(config.num_hidden_layers):
+            if self.hybrid_override_pattern[i] == "M":
+                # Mamba layer
+                self.conv_states += [
+                    torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
+                ]
+                self.ssm_states += [
+                    torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
+                ]
+            else:
+                # Attention or MLP layer
+                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.transformer_layers.append(i)
+
+        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Update the cache
+        if self.key_cache[layer_idx].shape[-1] == 0:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+        else:
+            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
+            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
+
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        for layer_idx in range(len(self.key_cache)):
+            device = self.key_cache[layer_idx].device
+            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.value_cache[layer_idx].device
+            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
+
+            device = self.conv_states[layer_idx].device
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.ssm_states[layer_idx].device
+            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
+        if len(self.key_cache) <= layer_idx:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    @classmethod
+    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    # Copied from modeling_mamba2.py
+    def update_conv_state(
+        self, layer_idx: int, new_conv_state: torch.Tensor, cache_init: bool = False
+    ) -> torch.Tensor:
+        if cache_init:
+            self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
+        else:
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
+            self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
+        return self.conv_states[layer_idx]
+
+    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
+        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
+        return self.ssm_states[layer_idx]
+
+    def reset(self):
+        self.conv_states.zero_()
+        self.ssm_states.zero_()
+
+class MambaRMSNormGated(torch.nn.Module):
+    def __init__(self, hidden_size, group_size, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+        self.group_size = group_size
+
+    # jan28b version
+    def forward(self, hidden_states, gate=None):
+        return rmsnorm_fn(x=hidden_states,
+                          weight=self.weight,
+                          bias=None, # No bias
+                          z=gate,
+                          eps=self.variance_epsilon,
+                          group_size=self.group_size,
+                          norm_before_gate=False
+        )
+
+class NemotronHMamba2Mixer(nn.Module):
+    """
+    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
+    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
+    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
+    and is why Mamba is called **selective** state spaces)
+    """
+
+    def __init__(self, config: NemotronHConfig, layer_idx: int):
+        super().__init__()
+        self.num_heads = config.mamba_num_heads
+        self.hidden_size = config.hidden_size
+        self.ssm_state_size = config.ssm_state_size
+        self.conv_kernel_size = config.conv_kernel
+        self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        self.layer_idx = layer_idx
+        self.use_conv_bias = config.use_conv_bias
+        self.activation = config.mamba_hidden_act
+        self.act = ACT2FN[config.mamba_hidden_act]
+
+        self.layer_norm_epsilon = config.layer_norm_epsilon
+
+        self.n_groups = config.n_groups
+        self.head_dim = config.mamba_head_dim
+        self.chunk_size = config.chunk_size
+
+        self.time_step_limit = config.time_step_limit
+        self.time_step_min = config.time_step_min
+        self.time_step_max = config.time_step_max
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=config.use_conv_bias,
+            kernel_size=config.conv_kernel,
+            groups=self.conv_dim,
+            padding=config.conv_kernel - 1,
+        )
+
+        # projection of the input hidden states
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size,
+            projection_size,
+            bias=config.use_bias,
+        )
+        # selective projection used to make dt, B and C input dependant
+
+        # time step projection (discretization)
+        # instantiate once and copy inv_dt in init_weights of PretrainedModel
+        self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
+
+        # S4D real initialization. These are not discretized!
+        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
+        A = torch.arange(1, self.num_heads + 1)
+        self.A_log = nn.Parameter(torch.log(A))
+        self.A_log._no_weight_decay = True
+        self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon, group_size=self.intermediate_size // self.n_groups)
+        self.D = nn.Parameter(torch.ones(self.num_heads))
+        self.D._no_weight_decay = True
+
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.use_bias = config.use_bias
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
+                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
+                " https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def cuda_kernels_forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        # 1. Gated MLP's linear projection
+        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+        projected_states = self.in_proj(hidden_states)
+
+        # Set up dimensions for reshapes later
+        batch_size, seq_len, _ = hidden_states.shape
+        groups_time_state_size = self.n_groups * self.ssm_state_size
+        d_mlp = (
+            projected_states.shape[-1]
+            - 2 * self.intermediate_size
+            - 2 * self.n_groups * self.ssm_state_size
+            - self.num_heads
+        ) // 2
+
+        # Single step calculations via cache
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
+                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+            )
+
+            # 2. Convolution sequence transformation
+            hidden_states_B_C = causal_conv1d_update(
+                hidden_states_B_C,
+                cache_params.conv_states[self.layer_idx],
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                self.activation,
+            )
+
+            hidden_states, B, C = torch.split(
+                hidden_states_B_C,
+                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                dim=-1,
+            )
+
+            # 3. SSM transformation
+            A = -torch.exp(self.A_log.float())  # (nheads,)
+            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
+            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
+            D = self.D[:, None, ...].expand(-1, self.head_dim)
+            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
+            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
+            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
+            hidden_states = selective_state_update(
+                cache_params.ssm_states[self.layer_idx],
+                hidden_states_reshaped,
+                dt,
+                A,
+                B,
+                C,
+                D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+            )
+            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
+            hidden_states = self.norm(hidden_states, gate)
+
+            # 4. Final linear projection
+            out = self.out_proj(hidden_states)[:, None, ...]
+
+        # Fused calculations or step by step if no initialized cache is found
+        else:
+            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
+            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
+
+            # 2-4. Fused kernel for conv1d, SSM, and the final projection
+            if self.training and cache_params is None:
+                out = mamba_split_conv1d_scan_combined(
+                    projected_states,
+                    self.conv1d.weight.squeeze(1),
+                    self.conv1d.bias,
+                    self.dt_bias,
+                    A,
+                    D=self.D,
+                    chunk_size=self.chunk_size,
+                    seq_idx=None,  # was seq_idx
+                    activation=self.activation,
+                    rmsnorm_weight=self.norm.weight,
+                    rmsnorm_eps=self.norm.variance_epsilon,
+                    outproj_weight=self.out_proj.weight,
+                    outproj_bias=self.out_proj.bias,
+                    headdim=self.head_dim,
+                    ngroups=self.n_groups,
+                    norm_before_gate=False,
+                    return_final_states=False,
+                    **dt_limit_kwargs,
+                )
+
+            else:
+                _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+                )
+
+                # 2. Convolution sequence transformation
+                # Init cache
+                if cache_params is not None:
+                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                    conv_states = nn.functional.pad(
+                        hidden_states_B_C_transposed,
+                        (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
+                    )
+                    cache_params.update_conv_state(
+                        layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
+                    )
+
+                if self.activation not in ["silu", "swish"]:
+                    hidden_states_B_C = self.act(
+                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
+                    )
+                else:
+                    hidden_states_B_C = causal_conv1d_fn(
+                        x=hidden_states_B_C.transpose(1, 2),
+                        weight=self.conv1d.weight.squeeze(1),
+                        bias=self.conv1d.bias,
+                        activation=self.activation,
+                    ).transpose(1, 2)
+                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+                hidden_states, B, C = torch.split(
+                    hidden_states_B_C,
+                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                    dim=-1,
+                )
+
+                # 3. SSM transformation
+                scan_output, ssm_state = mamba_chunk_scan_combined(
+                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
+                    dt,
+                    A,
+                    B.view(batch_size, seq_len, self.n_groups, -1),
+                    C.view(batch_size, seq_len, self.n_groups, -1),
+                    chunk_size=self.chunk_size,
+                    D=self.D,
+                    z=None,
+                    seq_idx=None,
+                    return_final_states=True,
+                    dt_bias=self.dt_bias,
+                    dt_softplus=True,
+                    **dt_limit_kwargs,
+                )
+
+                # Init cache
+                if ssm_state is not None and cache_params is not None:
+                    cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+                scan_output = scan_output.view(batch_size, seq_len, -1)
+
+                # Multiply "gate" branch and apply extra normalization layer
+                scan_output = self.norm(scan_output, gate)
+
+                # 4. Final linear projection
+                out = self.out_proj(scan_output)
+        return out
+
+    # fmt: off
+    def torch_forward(self, input_states, cache_params: Optional[HybridMambaAttentionDynamicCache]=None, cache_position:Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
+
+        # 1. Gated MLP's linear projection
+        input_states = apply_mask_to_padding_states(input_states, attention_mask)
+        projected_states = self.in_proj(input_states)
+        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
+        _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
+        )
+
+        # 2. Convolution sequence transformation
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
+
+            # We need to guarantee that anything regarding the cache is on the same device
+            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
+
+            hidden_states_B_C = torch.sum(
+                conv_states * self.conv1d.weight.squeeze(1), dim=-1
+            )
+            if self.use_conv_bias:
+                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
+            hidden_states_B_C = self.act(hidden_states_B_C)
+        else:
+            # Init cache
+            if cache_params is not None:
+                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                conv_states = nn.functional.pad(
+                    hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
+                )
+                cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
+
+            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))
+
+        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+        hidden_states, B, C = torch.split(
+            hidden_states_B_C,
+            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
+            dim=-1
+        )
+
+        # 3. SSM transformation
+        A = -torch.exp(self.A_log.float())                            # [num_heads]
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            # We need to guarantee that anything regarding the cache is on the same device
+            cache_device = cache_params.ssm_states.device
+
+            # Note: there is no need to pad parameter matrices here, as there is just one new token
+            # for batched generation
+            dt = dt[:, 0, :][:, None, ...]
+            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
+            # [num_heads] -> [num_heads, head_dim]
+            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
+
+            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            # [bsz, num_heads, head_dim, state_size]
+            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
+
+            # Discretize B
+            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
+            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
+            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
+            B = B.reshape(batch_size, -1, B.shape[-1])
+            # [bsz, num_heads, head_dim, state_size]
+            dB = dt[..., None] * B[..., None, :]
+
+            # Discretize x into dB
+            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
+            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
+            dBx = (dB * hidden_states[..., None]).to(device=cache_device)
+
+            # State calculation
+            cache_params.update_ssm_state(
+                layer_idx=self.layer_idx,
+                new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
+            )
+
+            # Subsequent output
+            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
+            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
+            C = C.reshape(batch_size, -1, C.shape[-1])
+            # [bsz, num_heads, head_dim]
+
+            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
+            # Reshape ssm_states to merge the first two dimensions
+            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
+            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
+            y = torch.bmm(ssm_states_reshaped, C_reshaped)
+            y = y.view(batch_size, self.num_heads, self.head_dim)
+
+            # D skip connection
+            # [num_heads] -> [num_heads, head_dim]
+            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
+            y = (y + hidden_states * D).to(y.dtype)
+
+            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
+            y = y.reshape(batch_size, -1)[:, None, ...]
+        else:
+            # begin ssd naive implementation without einsums
+            dt = nn.functional.softplus(dt + self.dt_bias)
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
+            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            B = B.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            C = C.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
+
+            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
+
+            # Discretize x and A
+            hidden_states = hidden_states * dt[..., None]
+            A = A.to(hidden_states.dtype) * dt
+
+            # Rearrange into blocks/chunks
+            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
+
+            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
+            A = A.permute(0, 3, 1, 2)
+            A_cumsum = torch.cumsum(A, dim=-1)
+
+            # 1. Compute the output for each intra-chunk (diagonal blocks)
+            # This is the analog of a causal mask
+            L = torch.exp(segment_sum(A))
+
+            # Contraction of C and B to get G (attention-weights like)
+            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]  # shape: (b, c, l, s, h, n)
+            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)
+
+            # Compute M, equivalent to applying attention mask to weights
+            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
+            M = M_intermediate.sum(dim=-1)
+
+            # Compute Y_diag (apply to values)
+            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
+
+            # 2. Compute the state for each intra-chunk
+            # (right term of low-rank factorization of off-diagonal blocks; B terms)
+            decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
+            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
+            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
+
+            # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
+            # (middle term of factorization of off-diag blocks; A terms)
+            if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
+            else:
+                previous_states = torch.zeros_like(states[:, :1])
+            states = torch.cat([previous_states, states], dim=1)
+            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
+            states, ssm_state = new_states[:, :-1], new_states[:, -1]
+
+            # 4. Compute state -> output conversion per chunk
+            # (left term of low-rank factorization of off-diagonal blocks; C terms)
+            state_decay_out = torch.exp(A_cumsum)
+            C_times_states = (C[..., None, :] * states[:, :, None, ...])
+            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
+            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
+
+            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
+            y = Y_diag + Y_off
+            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
+            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
+
+            y = y + D_residual
+            # Cutting off padded chunks
+            if pad_size > 0:
+                y = y[:, :seq_len, :, :]
+            y = y.reshape(batch_size, seq_len, -1)
+
+            # Init cache
+            if ssm_state is not None and cache_params is not None:
+                cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+        scan_output = self.norm(y, gate)
+
+        # end ssd naive
+
+        # 4. Final linear projection
+        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
+        return contextualized_states
+    # fmt: on
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
+            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
+        dtype = hidden_states.dtype
+        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
+            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)
+
+
+class NemotronHRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        NemotronHRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        # Weights are in float32
+        return (self.weight.to(torch.float32) * hidden_states).to(input_dtype)
+
+class NemotronHBlock(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.residual_in_fp32 = config.residual_in_fp32
+        self.norm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        # M: Mamba2, *: Attention, -: MLP
+        self.block_type = config.layers_block_type[layer_idx]
+        if self.block_type == "mamba":
+            self.mixer = NemotronHMamba2Mixer(config, layer_idx=layer_idx)
+        elif self.block_type == "attention":
+            self.mixer = NEMOTRONH_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+        elif self.block_type == "mlp":
+            self.mixer = NemotronHMLP(config, layer_idx=layer_idx)
+        else:
+            raise ValueError(f"Invalid layer pattern {config.hybrid_override_pattern[layer_idx]}")
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        with torch.cuda.stream(torch.cuda.default_stream(hidden_states.device)):
+            # * Use torch.cuda.stream() to avoid NaN issues when using multiple GPUs
+            residual = hidden_states
+            hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
+            if self.residual_in_fp32:
+                residual = residual.to(torch.float32)
+
+            if self.block_type == "mamba":
+                hidden_states = self.mixer(
+                    hidden_states, cache_params=cache_params, cache_position=cache_position
+                )
+            elif self.block_type == "attention":
+                hidden_states = self.mixer(
+                    hidden_states, cache_position=cache_position
+                )
+                hidden_states = hidden_states[0]
+            elif self.block_type == "mlp":
+                hidden_states = self.mixer(
+                    hidden_states
+                )
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            hidden_states = residual + hidden_states
+            return hidden_states
+
+
+# Copied from transformers.models.nemotron.modeling_nemotron Nemotron->NemotronH
+class NemotronHMLP(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+        self.hidden_size = config.hidden_size
+        #intermediate_size = config.expand * config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.mlp_hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.up_proj(x)))
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class NemotronHAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: NemotronHConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        if config.head_dim is not None:
+            self.head_dim = config.head_dim
+        else:
+            self.head_dim = config.hidden_size // config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.head_dim * self.num_heads, self.hidden_size, bias=config.attention_bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        # position_embeddings: Tuple[torch.Tensor, torch.Tensor], #TODO
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        #attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.num_heads * self.head_dim)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
+#class JambaFlashAttention2(JambaAttention):
+class NemotronHFlashAttention2(NemotronHAttention):
+    """
+    Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        #attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
+#class JambaSdpaAttention(JambaAttention):
+class NemotronHSdpaAttention(NemotronHAttention):
+    """
+    Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from NemotronHAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "NemotronHModel is using NemotronHSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+        is_causal = True if self.is_causal and causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+NEMOTRONH_ATTENTION_CLASSES = {
+    "eager": NemotronHAttention,
+    "flash_attention_2": NemotronHFlashAttention2,
+    "sdpa": NemotronHSdpaAttention,
+}
+
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2PreTrainedModel
+class NemotronHPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = NemotronHConfig
+    base_model_prefix = "backbone"
+    _no_split_modules = ["NemotronHBlock"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, NemotronHMamba2Mixer):
+            module.A_log._no_weight_decay = True
+            module.D._no_weight_decay = True
+
+            dt = torch.exp(
+                torch.rand(self.config.mamba_num_heads)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+
+            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                module.dt_bias.copy_(inv_dt)
+            module.dt_bias._no_reinit = True
+
+        if isinstance(module, nn.Linear):
+            if module.bias is not None:
+                if not getattr(module.bias, "_no_reinit", False):
+                    nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=self.config.initializer_range)
+
+        # TODO: Check
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            for name, p in module.named_parameters():
+                if name in ["out_proj.weight"]:
+                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                    # We need to reinit p since this code could be called multiple times
+                    # Having just p *= scale would repeatedly scale it down
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(self.config.num_hidden_layers)
+
+
+@dataclass
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2Output with MAMBA2->NemotronH,Mamba2->NemotronH
+class NemotronHOutput(ModelOutput):
+    """
+    Class for the NemotronH model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.mamba2.modeling_mamba2.MambaCausalLMOutput with Mamba2->NemotronH
+class NemotronHCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+NEMOTRONH_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`NemotronHConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+NEMOTRONH_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary.
+
+            If `cache_params.seqlen_offset>0`, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        position_ids (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        cache_params (`HybridMambaAttentionDynamicCache`, *optional*):
+            If passed along, the model uses the previous state in all the blocks (which will give the output for the
+            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
+        use_cache (`bool`, *optional*):
+            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            The position of the current input in the cache. This is used to ensure that the cache is correctly updated.
+            If `cache_params` is passed, `cache_position` should also be passed.
+        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+"""
+
+
+@add_start_docstrings(
+    "The bare NemotronH Model transformer outputting raw hidden-states without any specific head on top.",
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHModel(NemotronHPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([NemotronHBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        # Initialize weights and apply final processing
+        self._register_load_state_dict_pre_hook(self.load_hook)
+        self.post_init()
+
+    def load_hook(self, state_dict, prefix, *args):
+        for k in state_dict:
+            if "embedding." in k:
+                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
+                break
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, NemotronHOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        # use_cache = use_cache if use_cache is not None else self.config.use_cache
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # From zamba_modeling.py
+        if use_cache and cache_params is None:
+            logger.warning_once(
+                "NemotronH requires an initialized `NemotronHHybridDynamicCache` to return a cache. None was "
+                "provided, so no cache will be returned."
+            )
+
+        hidden_states = inputs_embeds
+
+        if cache_position is None:
+            cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
+        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        # Until HERE
+
+        for layer_idx, mixer_block in enumerate(self.layers):
+            # Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
+            if mixer_block.block_type == "mamba":
+                layer_mask = mamba_mask
+            elif mixer_block.block_type == "attention":
+                layer_mask = causal_mask
+            elif mixer_block.block_type == "mlp":
+                layer_mask = None
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__, hidden_states, cache_params, cache_position, layer_mask
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    cache_position=cache_position,
+                    attention_mask=layer_mask,
+                )
+
+            # TODO: Store attentions
+            # if output_attentions:
+            #     if layer_outputs[1] is not None:
+            #         # append attentions only of attention layers. Mamba layers return `None` as the attention weights
+            #         all_self_attns += (layer_outputs[1],)
+
+            # TODO (Check): should it happen before the forward pass?
+            # if output_hidden_states:
+            #     all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return NemotronHOutput(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    # Copied from transformers.models.jamba.modeling_jamba.JambaModel._update_causal_mask
+    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        target_length = cache_position[-1] + 1
+
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    def _update_mamba_mask(self, attention_mask, cache_position):
+        """
+        No need for zeroing states when
+            1. Cached forward
+            2. Attending to all inputs
+        """
+        mamba_mask = attention_mask
+        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
+            mamba_mask = None
+        return mamba_mask
+
+
+@add_start_docstrings(
+    """
+    The NEMOTRONH Model transformer with a language modeling head on top (linear layer with weights not tied to the input
+    embeddings).
+    """,
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHForCausalLM(NemotronHPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = NemotronHModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+        # Overwitten -- uses `cache_params` as opposed to `past_key_values`
+        empty_past_kv = past_key_values is None
+
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
+        #              (we can't check exception 3 while compiling)
+        if not empty_past_kv:
+            if (
+                inputs_embeds is not None  # Exception 1
+                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
+            ):
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+        else:
+            past_key_values = HybridMambaAttentionDynamicCache(
+                self.config, input_ids.shape[0], self.dtype, device=self.device
+            )
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if not empty_past_kv:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and empty_past_kv:
+            # TODO(pjin): workaround fix for properly extending inputs_embeds;
+            # longer term, may be better handled elsewhere in .generate().
+            if input_ids is not None and inputs_embeds.shape[1] < input_ids.shape[1]:
+                new_token_embeds = self.get_input_embeddings()(input_ids[:,inputs_embeds.shape[1]:])
+                inputs_embeds = torch.cat([inputs_embeds, new_token_embeds], dim=1)
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+                "logits_to_keep": self.config.num_logits_to_keep,
+                "cache_position": cache_position,
+            }
+        )
+        return model_inputs
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHCausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,  # for now we need this for generation
+    ) -> Union[Tuple, NemotronHCausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        nemotron_h_outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = nemotron_h_outputs[0]
+
+        # TODO: Check zamba_modeling.py: https://github.com/huggingface/transformers/blob/d7188ba600e36d3fd191b12e19f1b3bb81a8404f/src/transformers/models/zamba/modeling_zamba.py#L1284C1-L1286C2
+        #logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+        logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + nemotron_h_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return NemotronHCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=nemotron_h_outputs.cache_params,
+            hidden_states=nemotron_h_outputs.hidden_states,
+            attentions=nemotron_h_outputs.attentions,
+        )

NVIDIA-Nemotron-Nano-12B-v2/nemotron_toolcall_parser_no_streaming.py

@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import ast
+import json
+import re
+from collections.abc import Sequence
+from typing import Union
+
+import partial_json_parser
+from partial_json_parser.core.options import Allow
+
+from vllm.entrypoints.openai.protocol import (
+    ChatCompletionRequest,
+    DeltaFunctionCall, DeltaMessage,
+    DeltaToolCall,
+    ExtractedToolCallInformation,
+    FunctionCall,
+    ToolCall,
+)
+from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
+    ToolParser,
+    ToolParserManager,
+)
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.utils import random_uuid
+
+logger = init_logger(__name__)
+
+
+@ToolParserManager.register_module("nemotron_json")
+class NemotronJSONToolParser(ToolParser):
+
+    def __init__(self, tokenizer: AnyTokenizer):
+        super().__init__(tokenizer)
+
+        self.current_tool_name_sent: bool = False
+        self.prev_tool_call_arr: list[dict] = []
+        self.current_tool_id: int = -1
+        self.streamed_args_for_tool: list[str] = []
+
+        self.tool_call_start_token: str = "<TOOLCALL>"
+        self.tool_call_end_token: str = "</TOOLCALL>"
+
+        self.tool_call_regex = re.compile(r"<TOOLCALL>(.*?)</TOOLCALL>", re.DOTALL)
+
+    def extract_tool_calls(
+        self,
+        model_output: str,
+        request: ChatCompletionRequest,
+    ) -> ExtractedToolCallInformation:
+
+        if self.tool_call_start_token not in model_output:
+            return ExtractedToolCallInformation(
+                tools_called=False,
+                tool_calls=[],
+                content=model_output,
+            )
+
+        else:
+
+            try:
+                str_tool_calls = self.tool_call_regex.findall(model_output)[0].strip()
+                if not str_tool_calls.startswith("["):
+                    str_tool_calls = "[" + str_tool_calls
+                if not str_tool_calls.endswith("]"):
+                    str_tool_calls = "]" + str_tool_calls
+                json_tool_calls = json.loads(str_tool_calls)
+                tool_calls = []
+                for tool_call in json_tool_calls:
+                    try:
+                        tool_calls.append(ToolCall(
+                            type="function",
+                            function=FunctionCall(
+                                name=tool_call["name"],
+                                arguments=json.dumps(tool_call["arguments"], ensure_ascii=False) \
+                                    if isinstance(tool_call["arguments"], dict) else tool_call["arguments"],
+                            ),
+                        ))
+                    except:
+                        continue
+
+                content = model_output[:model_output.rfind(self.tool_call_start_token)]
+
+                return ExtractedToolCallInformation(
+                    tools_called=True,
+                    tool_calls=tool_calls,
+                    content=content if content else None,
+                )
+
+            except Exception:
+                logger.exception(f"Error in extracting tool call from response. Response: {model_output}")
+                return ExtractedToolCallInformation(
+                    tools_called=False,
+                    tool_calls=[],
+                    content=model_output,
+                )
+
+    def extract_tool_calls_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+        request: ChatCompletionRequest,
+    ) -> Union[DeltaMessage, None]:
+
+        raise NotImplementedError("Tool calling is not supported in streaming mode!")

NVIDIA-Nemotron-Nano-12B-v2/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<SPECIAL_12>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

NVIDIA-Nemotron-Nano-12B-v2/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3277c00fe5fb3963b3cb7c07b7f183722d2af4d775a4aea7cfb3684d7cccbc2f
+size 17078330

NVIDIA-Nemotron-Nano-31B-A3-v3/.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text