Skip to content

[Spec] Add trtllm_mha support for Gemma 4 MTP draft attention backend #25545

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kpham-sgl wants to merge 1 commit into
base: kp/gemma4-mtp-remove-seed
Choose a base branch
from
+23 −6
Open

[Spec] Add trtllm_mha support for Gemma 4 MTP draft attention backend #25545

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 23 additions & 6 deletions python/sglang/srt/speculative/frozen_kv_mtp_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,12 +205,14 @@ def _resolve_draft_backend_type(self) -> str:

def _init_draft_attn_backend(self):
backend_type = self._resolve_draft_backend_type()
if backend_type != "triton":
raise ValueError(
"Frozen-KV MTP currently supports only the triton attention "
f"backend, got {backend_type}."
)
return self._init_triton_draft_attn_backend()
if backend_type == "triton":
return self._init_triton_draft_attn_backend()
if backend_type == "trtllm_mha":
return self._init_trtllm_mha_draft_attn_backend()
raise ValueError(
"Frozen-KV MTP currently supports triton and trtllm_mha attention "
f"backends, got {backend_type}."
)

def _init_triton_draft_attn_backend(self):
from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
Expand All @@ -225,6 +227,21 @@ def _init_triton_draft_attn_backend(self):
kv_indptr_buf=kv_indptr_buf,
)

def _init_trtllm_mha_draft_attn_backend(self):
# TODO(kpham-sgl): trtllm_mha (Gemma4) known gaps:
# 1. target=trtllm_mha + num_draft_tokens>=6: flashinfer missing
# tile (tileSizeQ=64, headDim=512, page=64) for target_verify
# on full-attn layers.
# 2. target=triton + draft=trtllm_mha: needs --page-size in
# {16,32,64} (triton defaults to 1; no auto-promote for
# draft-only trtllm_mha).
# 3. topk>1 + page_size>1 + SWA: eagle verify hits
# `move_kv_cache` w/o enable_kv_cache_copy (orthogonal to
# trtllm_mha but blocks the only triton-target workaround).
from sglang.srt.layers.attention.trtllm_mha_backend import TRTLLMHAAttnBackend

return TRTLLMHAAttnBackend(self.draft_model_runner, skip_prefill=True)

def _bind_kv_context(self) -> None:
draft_model = self.draft_model_runner.model
if not hasattr(draft_model, "build_frozen_kv_mtp_context") or not hasattr(
Expand Down
Loading