Document LA Flash implementation details

kernel_utils/README.md

@@ -25,6 +25,31 @@ types:
 | `0` | Full attention over the listed key segment or packed key segments. |
 | `1` | Bottom-right causal attention. |
 
+## How It Works
+
+`batch_utils.hybrid_runtime` builds sparse range plans for the text decoder.
+Each plan describes which query token intervals attend to which key/value token
+intervals. `kernel_utils.range_attention` executes those plans with
+FlashAttention instead of materializing dense SDPA masks.
+
+The runtime follows three paths:
+
+- **Packed simple plans:** when each query range maps to one contiguous
+  key/value range, LA Flash flattens the selected ranges, builds FlashAttention
+  `cu_seqlens_q` / `cu_seqlens_k`, and calls `flash_attn_varlen_func` directly.
+- **Packed MTP full-window plans:** for hybrid MTP decode, multiple full
+  key/value windows for the same query block are concatenated into one packed
+  key/value sequence before the FlashAttention call. This keeps the sparse
+  memory profile without constructing a `[B,H,Q,K]` attention mask.
+- **Compatible multi-segment plans:** when a query range attends to multiple
+  segments that cannot be packed as one sequence, each segment is evaluated with
+  FlashAttention and the partial outputs are merged with the standard
+  log-sum-exp softmax composition.
+
+The output tensor shape and dtype match the decoder attention output expected
+by the model. This path is inference-oriented and depends on FlashAttention's
+forward kernels; it is not a custom autograd training backend.
+
 ## Runtime Knobs
 
 | Variable | Default | Meaning |