Upload folder using huggingface_hub

README.md

@@ -1,3 +1,14 @@
+---
+title: PowerPoint Template Customizer
+emoji: "📊"
+colorFrom: blue
+colorTo: gray
+sdk: gradio
+sdk_version: 5.34.2
+app_file: app.py
+pinned: false
+---
+
 # PowerPoint Template Customizer
 
 Generate styled presentations from a user-provided `.pptx` design template, guided by LLM-produced structured slide JSON.
@@ -33,10 +44,15 @@ python app.py
 
 ## Backend Options
 Set `MODEL_BACKEND` environment variable:
-- `transformers` (default): uses `unsloth/gemma-4-12b-it` through Hugging Face Transformers.
-- `llama.cpp`: uses local GGUF via `LLAMA_CPP_MODEL_PATH`.
+- `llama.cpp` (default): uses GGUF and auto-resolves from `unsloth/gemma-4-12b-it-GGUF` when `LLAMA_CPP_MODEL_PATH` is not set.
+- `transformers`: optional fallback that uses `unsloth/gemma-4-12b-it` through Hugging Face Transformers.
 - `vllm`: uses OpenAI-compatible endpoint (`VLLM_BASE_URL`, `VLLM_MODEL_NAME`).
 
+GGUF-related environment variables:
+- `GGUF_MODEL_REPO` (default: `unsloth/gemma-4-12b-it-GGUF`)
+- `GGUF_CACHE_DIR` (default: `models`)
+- `LLAMA_CPP_MODEL_PATH` (optional explicit path override)
+
 ## Quick Local Backend Smoke Test (`llama.cpp`)
 ```bash
 pip install llama-cpp-python huggingface_hub openai
@@ -50,6 +66,13 @@ LLAMA_CPP_MODEL_PATH=/absolute/path/to/model.gguf \
 python scripts/local_backend_smoke.py
 ```
 
+Or rely on auto-download from Hugging Face:
+```bash
+MODEL_BACKEND=llama.cpp \
+GGUF_MODEL_REPO=unsloth/gemma-4-12b-it-GGUF \
+python scripts/local_backend_smoke.py
+```
+
 ## Tests
 ```bash
 pytest -q

requirements.txt

@@ -3,4 +3,7 @@ transformers
 torch
 accelerate
 python-pptx
-spaces
+spaces
+llama-cpp-python
+huggingface_hub
+openai

scripts/local_backend_smoke.py

@@ -19,9 +19,15 @@ def main() -> int:
     os.environ.setdefault("MODEL_BACKEND", backend)
 
     if backend == "llama.cpp" and not os.getenv("LLAMA_CPP_MODEL_PATH"):
-        candidate = repo_root / "models" / "Qwen2.5-0.5B-Instruct-Q4_K_M.gguf"
-        if candidate.exists():
-            os.environ["LLAMA_CPP_MODEL_PATH"] = str(candidate)
+        candidates = [
+            repo_root / "models" / "gemma-4-12b-it-Q4_K_M.gguf",
+            repo_root / "models" / "Gemma-4-12b-it-Q4_K_M.gguf",
+            repo_root / "models" / "Qwen2.5-0.5B-Instruct-Q4_K_M.gguf",
+        ]
+        for candidate in candidates:
+            if candidate.exists():
+                os.environ["LLAMA_CPP_MODEL_PATH"] = str(candidate)
+                break
 
     output = customize_presentation(
         str(template),

src/llm/gemma_client.py

@@ -1,8 +1,8 @@
 from __future__ import annotations
 
-import json
 import os
 from dataclasses import dataclass
+from pathlib import Path
 
 import spaces
 import torch
@@ -22,10 +22,11 @@ SYSTEM_PROMPT = (
 
 @dataclass
 class GemmaClientConfig:
-    model_id: str = "unsloth/gemma-4-12b-it"
+    model_id: str = "unsloth/gemma-4-12b-it-GGUF"
+    transformers_model_id: str = "unsloth/gemma-4-12b-it"
     temperature: float = 0.4
     max_new_tokens: int = 900
-    backend: str = "transformers"  # transformers | llama.cpp | vllm
+    backend: str = "llama.cpp"  # llama.cpp | transformers | vllm
 
 
 class GemmaClient:
@@ -38,7 +39,7 @@ class GemmaClient:
         if self._model is not None and self._tokenizer is not None:
             return
 
-        model_id = self.config.model_id
+        model_id = self.config.transformers_model_id
         self._tokenizer = AutoTokenizer.from_pretrained(model_id)
 
         kwargs = {
@@ -98,9 +99,7 @@ class GemmaClient:
                 "llama.cpp backend requested but llama-cpp-python is not installed."
             ) from exc
 
-        model_path = os.getenv("LLAMA_CPP_MODEL_PATH")
-        if not model_path:
-            raise RuntimeError("Set LLAMA_CPP_MODEL_PATH to a local GGUF file for llama.cpp backend.")
+        model_path = self._resolve_llama_cpp_model_path()
 
         llm = Llama(
             model_path=model_path,
@@ -118,6 +117,52 @@ class GemmaClient:
         )
         return completion["choices"][0]["message"]["content"].strip()
 
+    def _resolve_llama_cpp_model_path(self) -> str:
+        env_model_path = os.getenv("LLAMA_CPP_MODEL_PATH")
+        if env_model_path:
+            return env_model_path
+
+        preferred = [
+            "Q4_K_M.gguf",
+            "q4_k_m.gguf",
+            "Q4_K_S.gguf",
+            "q4_k_s.gguf",
+            "Q5_K_M.gguf",
+            "q5_k_m.gguf",
+        ]
+
+        try:
+            from huggingface_hub import HfApi, hf_hub_download
+        except Exception as exc:
+            raise RuntimeError(
+                "llama.cpp backend requires either LLAMA_CPP_MODEL_PATH or huggingface_hub to auto-download GGUF."
+            ) from exc
+
+        repo_id = os.getenv("GGUF_MODEL_REPO", self.config.model_id)
+        files = HfApi(token=os.getenv("HF_TOKEN")).list_repo_files(repo_id=repo_id)
+        gguf_files = [f for f in files if f.lower().endswith(".gguf")]
+        if not gguf_files:
+            raise RuntimeError(f"No GGUF files found in repo: {repo_id}")
+
+        selected = None
+        for suffix in preferred:
+            selected = next((f for f in gguf_files if f.endswith(suffix)), None)
+            if selected:
+                break
+        if selected is None:
+            selected = gguf_files[0]
+
+        local_dir = Path(os.getenv("GGUF_CACHE_DIR", "models"))
+        local_dir.mkdir(parents=True, exist_ok=True)
+        downloaded = hf_hub_download(
+            repo_id=repo_id,
+            filename=selected,
+            local_dir=str(local_dir),
+            token=os.getenv("HF_TOKEN"),
+        )
+        os.environ["LLAMA_CPP_MODEL_PATH"] = downloaded
+        return downloaded
+
     def _generate_vllm(self, user_prompt: str) -> str:
         try:
             from openai import OpenAI
@@ -159,7 +204,7 @@ _DEFAULT_CLIENT: GemmaClient | None = None
 def load_model(backend: str | None = None) -> GemmaClient:
     global _DEFAULT_CLIENT
     if _DEFAULT_CLIENT is None or (backend and _DEFAULT_CLIENT.config.backend != backend):
-        cfg = GemmaClientConfig(backend=backend or os.getenv("MODEL_BACKEND", "transformers"))
+        cfg = GemmaClientConfig(backend=backend or os.getenv("MODEL_BACKEND", "llama.cpp"))
         _DEFAULT_CLIENT = GemmaClient(cfg)
     return _DEFAULT_CLIENT
 

src/services/customization_service.py

@@ -31,7 +31,7 @@ def customize_presentation(template_input, user_input: str) -> str:
     template_path = _resolve_template_path(template_input)
 
     style_profile = parse_template_style(template_path)
-    backend = os.getenv("MODEL_BACKEND", "transformers")
+    backend = os.getenv("MODEL_BACKEND", "llama.cpp")
     slides_json = generate_slides_json(user_input.strip(), backend=backend)
 
     output_path = compile_presentation(