A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 22:45 UTC
[Refactor] Remove dead private func `_fp8_perm` and `_extract_mask_for_item` (#35068)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Feb 23, 2026 at 08:05 UTC
7f40e9e5164af6af34bcbd5945356d416e29e71a
2 files changed
-46
vllm/model_executor/layers/fused_moe/utils.py
-10
@@ -284,16 +284,6 @@ def moe_kernel_quantize_input(
284
return A, A_scale
285
286
287
-def _fp8_perm(m: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
288
- """
289
- A permutation routine that works on fp8 types.
290
- """
291
- if torch.is_floating_point(m) and m.dtype.itemsize == 1:
292
- return m.view(dtype=torch.uint8)[idx, ...].view(dtype=m.dtype)
293
- else:
294
- return m[idx, ...]
295
-
296
-
287
def normalize_scales_shape(scales: torch.Tensor | None) -> torch.Tensor | None:
288
if scales is not None:
289
if scales.numel() == 1:
vllm/model_executor/models/glmasr_utils.py
-36
@@ -130,39 +130,3 @@ def _group_audio_embeddings(
130
grouped_embeddings.append(torch.cat(audio_chunks, dim=0))
131
current_idx += count
132
return tuple(grouped_embeddings)
133
-
134
-
135
-def _normalize_to_tensor(mask: torch.Tensor | list[torch.Tensor]) -> torch.Tensor:
136
- """Convert mask to tensor, handling both list and tensor formats."""
137
- if isinstance(mask, list):
138
- return (
139
- torch.stack(mask)
140
- if mask and isinstance(mask[0], torch.Tensor)
141
- else torch.tensor(mask)
142
- )
143
- return mask
144
-
145
-
146
-def _extract_mask_for_item(
147
- feature_attention_mask: torch.Tensor | list[torch.Tensor],
148
- chunk_counts: torch.Tensor | list[int] | None,
149
- item_idx: int,
150
-) -> torch.Tensor:
151
- """Extract attention mask for a specific audio item."""
152
- if chunk_counts is None:
153
- # Single item per audio
154
- mask = feature_attention_mask[item_idx]
155
- if isinstance(feature_attention_mask, torch.Tensor):
156
- return mask.unsqueeze(0)
157
- return _normalize_to_tensor(mask)
158
-
159
- # Multiple chunks per audio: calculate slice indices
160
- counts = _as_list_chunk_counts(chunk_counts)
161
- start_idx = sum(counts[:item_idx])
162
- end_idx = start_idx + counts[item_idx]
163
-
164
- # Extract slice
165
- if isinstance(feature_attention_mask, torch.Tensor):
166
- return feature_attention_mask[start_idx:end_idx]
167
- mask_slice = feature_attention_mask[start_idx:end_idx]
168
- return _normalize_to_tensor(mask_slice)