From 5678b71b6ccaaaa5d586d2a9962e4f90c46247ce Mon Sep 17 00:00:00 2001 From: paulzhang-tm Date: Sun, 11 Oct 2026 01:39:17 -0400 Subject: [PATCH] [Spec] Allow Qwen3.5 backends to opt into packed target verification (#43037) --- python/sglang/srt/models/qwen3_5.py | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-) diff --git a/python/sglang/srt/models/qwen3_5.py b/python/sglang/srt/models/qwen3_5.py index 5ff875082a..def6a402d5 100644 --- a/python/sglang/srt/models/qwen3_5.py +++ b/python/sglang/srt/models/qwen3_5.py @@ -998,16 +998,23 @@ class Qwen3_5GatedDeltaNet(nn.Module): ) use_strided_prefill_z = False - use_fused_decode_proj_conv = ( + packed_verify = False + if forward_batch.forward_mode.is_target_verify(): + from sglang.srt.model_executor.forward_context import get_attn_backend + + backend = get_attn_backend() + backend = getattr(backend, "linear_attn_backend", backend) + packed_verify = getattr(backend, "supports_packed_verify", False) + use_fused_proj_conv = ( _gdn_decode_fused_proj_conv - and forward_batch.forward_mode.is_decode() + and (forward_batch.forward_mode.is_decode() or packed_verify) and isinstance(projected_states_qkvz, torch.Tensor) and isinstance(projected_states_ba, torch.Tensor) ) use_fused_contiguous_unpack = ( self.num_v_heads // self.num_k_heads in _GDN_FUSED_QKVZBA_RATIOS ) - if use_fused_decode_proj_conv: + if use_fused_proj_conv: # GDN owns indexed Conv1D state and the safe unpack/Conv boundary; # it replaces these temporary B/A placeholders before recurrence. mixed_qkv = (projected_states_qkvz, projected_states_ba) @@ -1055,11 +1062,10 @@ class Qwen3_5GatedDeltaNet(nn.Module): a=a, b=b, ) - if use_fused_decode_proj_conv: + if use_fused_proj_conv: if not isinstance(attn_result, tuple) or len(attn_result) != 2: raise RuntimeError( - "Fused GDN decode projection/Conv1D backend must return " - "(core_attn_out, z)" + "Fused GDN projection/Conv1D backend must return (core_attn_out, z)" ) core_attn_out, z = attn_result else: