EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Revert "[Fix] DeepSeek EP accuracy issue on B200 GPUs (#9946)" (#9955)

Browse Source
This commit is contained in:
Yineng Zhang
2025-09-02 23:49:56 -07:00
committed by GitHub
parent df397a72e8
commit 2c7ca33abb
3 changed files with 1 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/layers/quantization/deep_gemm_wrapper/compile_utils.py
View File

@@ -44,13 +44,6 @@ def update_deep_gemm_config(gpu_id: int, server_args: ServerArgs):
global _DO_COMPILE_ALL global _DO_COMPILE_ALL
global _IS_FIRST_RANK_ON_NODE global _IS_FIRST_RANK_ON_NODE
# Update UE8M0 scaling configuration based on server args
from sglang.srt.layers.quantization.deep_gemm_wrapper.configurer import (
update_deepgemm_scale_ue8m0,
)
update_deepgemm_scale_ue8m0(server_args.disable_deepgemm_ue8m0)
# Generate m_max # Generate m_max
m_max = 1024 * 16 m_max = 1024 * 16
if server_args.chunked_prefill_size < 1: if server_args.chunked_prefill_size < 1:

19
python/sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py
View File

@@ -29,21 +29,4 @@ def _is_blackwell_arch() -> bool:
ENABLE_JIT_DEEPGEMM = _compute_enable_deep_gemm() ENABLE_JIT_DEEPGEMM = _compute_enable_deep_gemm()
DEEPGEMM_BLACKWELL = ENABLE_JIT_DEEPGEMM and _is_blackwell_arch() DEEPGEMM_BLACKWELL = ENABLE_JIT_DEEPGEMM and _is_blackwell_arch()
# Allow disabling UE8M0 scaling for accuracy-critical workloads DEEPGEMM_SCALE_UE8M0 = DEEPGEMM_BLACKWELL
# This can help with DeepSeek EP accuracy issues on B200 GPUs
# Will be updated by server args in update_deepgemm_scale_ue8m0()
DEEPGEMM_SCALE_UE8M0 = DEEPGEMM_BLACKWELL and get_bool_env_var(
"SGL_ENABLE_DEEPGEMM_UE8M0", default="true"
)
def update_deepgemm_scale_ue8m0(disable_ue8m0: bool):
"""Update DEEPGEMM_SCALE_UE8M0 based on server arguments."""
global DEEPGEMM_SCALE_UE8M0
if disable_ue8m0:
DEEPGEMM_SCALE_UE8M0 = False
logger.info("DeepGEMM UE8M0 scaling disabled via server argument")
else:
DEEPGEMM_SCALE_UE8M0 = DEEPGEMM_BLACKWELL and get_bool_env_var(
"SGL_ENABLE_DEEPGEMM_UE8M0", default="true"
)

6
python/sglang/srt/server_args.py
View File

@@ -268,7 +268,6 @@ class ServerArgs:
flashinfer_mxfp4_moe_precision: Literal["default", "bf16"] = "default" flashinfer_mxfp4_moe_precision: Literal["default", "bf16"] = "default"
enable_flashinfer_allreduce_fusion: bool = False enable_flashinfer_allreduce_fusion: bool = False
deepep_mode: Literal["auto", "normal", "low_latency"] = "auto" deepep_mode: Literal["auto", "normal", "low_latency"] = "auto"
disable_deepgemm_ue8m0: bool = False
ep_num_redundant_experts: int = 0 ep_num_redundant_experts: int = 0
ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None
init_expert_location: str = "trivial" init_expert_location: str = "trivial"
@@ -1563,11 +1562,6 @@ class ServerArgs:
default="auto", default="auto",
help="Select the mode when enable DeepEP MoE, could be `normal`, `low_latency` or `auto`. Default is `auto`, which means `low_latency` for decode batch and `normal` for prefill batch.", help="Select the mode when enable DeepEP MoE, could be `normal`, `low_latency` or `auto`. Default is `auto`, which means `low_latency` for decode batch and `normal` for prefill batch.",
) )
parser.add_argument(
"--disable-deepgemm-ue8m0",
action="store_true",
help="Disable DeepGEMM UE8M0 scaling optimizations. This can help with accuracy issues on Blackwell GPUs (B200) for certain models like DeepSeek.",
)
parser.add_argument( parser.add_argument(
"--ep-num-redundant-experts", "--ep-num-redundant-experts",
type=int, type=int,