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

[Performance] Qwen3-Next: speed up update_mamba_state_after_mtp_verify by 10x; e2e up to 3.54% faster (#10586)

Browse Source
This commit is contained in:
Binyao Jiang
2025-09-18 17:13:59 -07:00
committed by GitHub
parent b91cb67e7a
commit 4e6c4923a0
1 changed files with 10 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
View File

@@ -583,36 +583,15 @@ class HybridLinearAttnBackend(AttentionBackend):
# Compute common indices once to avoid duplication # Compute common indices once to avoid duplication
last_steps_all = (accepted_length - 1).to(torch.int64) last_steps_all = (accepted_length - 1).to(torch.int64)
valid_state_indices = state_indices_tensor[valid_mask].to(torch.int64) valid_state_indices = state_indices_tensor[valid_mask].to(torch.int64) # [N]
last_steps = last_steps_all[valid_mask].to(torch.int64) last_steps = last_steps_all[valid_mask].to(torch.int64) # [N]
if valid_state_indices.numel() > 0: # scatter into ssm_states at the chosen cache lines
chunk = 256 ssm_states[:, valid_state_indices, :] = intermediate_state_cache[
num_valid = valid_state_indices.numel() :, valid_state_indices, last_steps
].to(ssm_states.dtype, copy=False)
# SSM state updates # Scatter into conv_states at the chosen cache lines
for i in range(0, num_valid, chunk): conv_states[:, valid_state_indices, :, :] = intermediate_conv_window_cache[
idx = valid_state_indices[i : i + chunk] :, valid_state_indices, last_steps
steps = last_steps[i : i + chunk] ].to(conv_states.dtype, copy=False)
# per (cache line, step)
for j in range(idx.numel()):
ci = idx[j].item()
st = steps[j].item()
ssm_states[:, ci, :].copy_(
intermediate_state_cache[:, ci, st].to(
ssm_states.dtype, copy=False
)
)
# Conv window updates
for i in range(0, num_valid, chunk):
idx = valid_state_indices[i : i + chunk]
steps = last_steps[i : i + chunk]
for j in range(idx.numel()):
ci = idx[j].item()
st = steps[j].item()
conv_states[:, ci, :, :].copy_(
intermediate_conv_window_cache[:, ci, st].to(
conv_states.dtype, copy=False
)
)