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

2026-06-15 09:44:21 -04:00
parent 09e9c4fde3
commit d5899b95c4
1 changed files with 12 additions and 4 deletions
@@ -119,10 +119,10 @@ _is_gfx95 = is_gfx95_supported()
 _is_hip = 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
-_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
 )
-_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
 )
 _is_amx_available = cpu_has_amx_support()
@@ -594,7 +594,11 @@ class Qwen3_5LinearDecoderLayer(nn.Module):
                layer_id=layer_id,
                config=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", "")),
                is_nextn=is_nextn,
                support_shared_expert_fusion=not _disable_shared_experts_fusion(),
@@ -806,7 +810,11 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
                layer_id=layer_id,
                config=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", "")),
                is_nextn=is_nextn,
                support_shared_expert_fusion=not _disable_shared_experts_fusion(),