chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Revert "fix llama4 kv cache layout" (#12437)

Browse Source
This commit is contained in:
b8zhong
2025-10-30 22:33:37 -07:00
committed by GitHub
parent 72b5f3d0bb
commit a076ec1a7a
2 changed files with 1 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/server_args.py
View File

@@ -980,13 +980,6 @@ class ServerArgs:
logger.warning(
"Use trtllm_mha as attention backend on sm100 for Llama4 model"
)
if is_sm100_supported() and self.attention_backend == "trtllm_mha":
# TODO(brayden): remove this once TRTLLM MHA kernel for FP8 w/ tileSizeKv=128 is available.
# This is a Llama 4 specific issue only.
self.kv_cache_dtype = "bfloat16"
logger.warning(
"Setting kv_cache_dtype to bfloat16 for Llama4 with trtllm_mha backend, due to a missing FlashInfer TRTLLM MHA kernel for FP8 KV Cache"
)
if is_sm100_supported() and self.moe_runner_backend == "auto":
if self.quantization in {"fp8", "modelopt_fp8"}:
self.moe_runner_backend = "flashinfer_trtllm"