xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-09 20:07:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix chunked prefill for GGUF (#14666)

Browse Source 
Signed-off-by: SzymonOzog <szymon.ozog@aleph-alpha.com>
This commit is contained in:
Szymon Ożóg 2025-03-13 08:19:03 +01:00 committed by GitHub
parent 5d043c1685
commit 55211b01e8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/layers/quantization/gguf.py
View File

@ -98,6 +98,13 @@ MMQ_QUANT_TYPES = STANDARD_QUANT_TYPES | KQUANT_TYPES
def _fuse_mul_mat(x: torch.Tensor, qweight: torch.Tensor, def _fuse_mul_mat(x: torch.Tensor, qweight: torch.Tensor,
qweight_type: int) -> torch.Tensor: qweight_type: int) -> torch.Tensor:
# HACK: when doing chunked prefill we don't generate output tokens
# so input to logits generator is empty which causes invalid parameter
if x.shape[0] == 0:
return torch.empty(x.shape[0],
qweight.shape[0],
dtype=x.dtype,
device=x.device)
# there is no need to call any kernel for fp16/bf16 # there is no need to call any kernel for fp16/bf16
if qweight_type in UNQUANTIZED_TYPES: if qweight_type in UNQUANTIZED_TYPES:
return x @ qweight.T return x @ qweight.T