xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-25 04:07:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

[bugfix] Fix Llama3/4 issues caused by FlashInfer 0.2.10 (#22426)

Browse Source 
Signed-off-by: Po-Han Huang <pohanh@nvidia.com>
This commit is contained in:
Po-Han Huang (NVIDIA) 2025-08-08 11:25:01 +08:00 committed by GitHub
parent 157f9c1368
commit af473f0a85
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 17 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
View File

@ -6,14 +6,22 @@ import torch
def calculate_tile_tokens_dim(num_tokens, top_k, num_experts):
from flashinfer import next_positive_power_of_2
# Guess tokens per expert assuming perfect expert distribution first.
num_tokens_per_expert = (num_tokens * top_k) // num_experts
# And pad the number to the next power of 2.
tile_tokens_dim = next_positive_power_of_2(num_tokens_per_expert)
# Cap to 8-64 tokens per CTA tile as it's the range supported by the kernel.
tile_tokens_dim = min(max(tile_tokens_dim, 8), 64)
# FlashInfer 0.2.10 has issues with larger tile sizes. Set to 8 for now.
# TODO: Revert this to dynamic calculation once a new version of FlashInfer
# with the necessary kernels is released.
tile_tokens_dim = 8
# from flashinfer import next_positive_power_of_2
# # Guess tokens per expert assuming perfect expert distribution first.
# num_tokens_per_expert = (num_tokens * top_k) // num_experts
# # And pad the number to the next power of 2.
# tile_tokens_dim = next_positive_power_of_2(num_tokens_per_expert)
# # Cap to 8-64 tokens per CTA tile as it's the range supported by the
# # kernel.
# tile_tokens_dim = min(max(tile_tokens_dim, 8), 64)
return tile_tokens_dim

3
vllm/v1/attention/backends/flashinfer.py
View File

@ -524,7 +524,8 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
head_dim = self.kv_cache_spec.head_size
# currently prefill trtllm attention does not support fp8 kv cache
prefill_use_trtllm = use_trtllm_attention(
prefill_use_trtllm = not cache_dtype.startswith("fp8") \
and use_trtllm_attention(
num_prefill_tokens, max_seq_len, cache_dtype,
num_qo_heads, num_kv_heads, head_dim)
decode_use_trtllm = use_trtllm_attention(