xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-11 05:05:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bug] Fix Qwen3-VL 2:4 sparsity shape mismatch during decompression

Browse Source 
Use the shape tensor loaded from the model checkpoint instead of
computing from layer attributes (logical_widths, input_size_per_partition)
which may not be properly initialized for some layer types like
RowParallelLinear in vision models.

Fixes #31019

Signed-off-by: yurekami <yurekami@users.noreply.github.com>

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
yurekami 2025-12-24 04:03:23 +09:00
parent c016c95b45
commit 8625dee2c0
1 changed files with 6 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py
View File

@ -379,13 +379,15 @@ class CompressedTensors24(CompressedTensorsScheme):
] ]
decompressed = combine_shards(decompressed_shards) decompressed = combine_shards(decompressed_shards)
else: else:
# Use the shape loaded from the model checkpoint rather than
# computing from layer attributes, as some layer types (e.g.,
# RowParallelLinear in vision models) may not have logical_widths
# or input_size_per_partition properly initialized.
out_dim, in_dim = layer.shape.data.view(-1).tolist()
decompressed = sparsity_compressor.decompress_weight( decompressed = sparsity_compressor.decompress_weight(
dict( dict(
compressed=compressed, compressed=compressed,
shape=( shape=(out_dim, in_dim),
layer.logical_widths[0],
layer.input_size_per_partition,
),
bitmask=bitmask, bitmask=bitmask,
) )
) )