minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix(qwen3.5): keep CUDA dual-stream overlap (regressed by #25885) (#27868)

Browse Source
This commit is contained in:
YAMY
2026-06-15 09:44:21 -04:00
committed by GitHub
parent 09e9c4fde3
commit d5899b95c4
1 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_5.py
+12 -4
View File
@@ -119,10 +119,10 @@ _is_gfx95 = is_gfx95_supported()
_is_hip = is_hip() _is_hip = is_hip()
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
_hip_use_alt_stream = get_bool_env_var("SGLANG_ALT_STREAM") and _is_hip _hip_use_alt_stream = get_bool_env_var("SGLANG_ALT_STREAM") and _is_hip
_gdn_use_alt_stream = ( _gdn_use_alt_stream = _is_cuda or (
get_bool_env_var("SGLANG_GDN_QKVZ_BA_ALT_STREAM", "False") and _hip_use_alt_stream get_bool_env_var("SGLANG_GDN_QKVZ_BA_ALT_STREAM", "False") and _hip_use_alt_stream
) )
_qknorm_use_alt_stream = ( _qknorm_use_alt_stream = _is_cuda or (
get_bool_env_var("SGLANG_QK_NORM_ALT_STREAM", "False") and _hip_use_alt_stream get_bool_env_var("SGLANG_QK_NORM_ALT_STREAM", "False") and _hip_use_alt_stream
) )
_is_amx_available = cpu_has_amx_support() _is_amx_available = cpu_has_amx_support()
@@ -594,7 +594,11 @@ class Qwen3_5LinearDecoderLayer(nn.Module):
layer_id=layer_id, layer_id=layer_id,
config=config, config=config,
quant_config=quant_config, quant_config=quant_config,
alt_stream=(alt_stream if _disable_shared_experts_fusion() else None), alt_stream=(
alt_stream
if (_is_cuda or _disable_shared_experts_fusion())
else None
),
prefix=add_prefix("mlp", prefix.replace(".linear_attn", "")), prefix=add_prefix("mlp", prefix.replace(".linear_attn", "")),
is_nextn=is_nextn, is_nextn=is_nextn,
support_shared_expert_fusion=not _disable_shared_experts_fusion(), support_shared_expert_fusion=not _disable_shared_experts_fusion(),
@@ -806,7 +810,11 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
layer_id=layer_id, layer_id=layer_id,
config=config, config=config,
quant_config=quant_config, quant_config=quant_config,
alt_stream=(alt_stream if _disable_shared_experts_fusion() else None), alt_stream=(
alt_stream
if (_is_cuda or _disable_shared_experts_fusion())
else None
),
prefix=add_prefix("mlp", prefix.replace(".self_attn", "")), prefix=add_prefix("mlp", prefix.replace(".self_attn", "")),
is_nextn=is_nextn, is_nextn=is_nextn,
support_shared_expert_fusion=not _disable_shared_experts_fusion(), support_shared_expert_fusion=not _disable_shared_experts_fusion(),