Skip to content

[https://nvbugs/6191524][fix] In MLA.forward_context, also call the warmup when has_cached_kv_for_mla_context #14536

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

Merged
byshiue merged 2 commits into from
Jun 2, 2026
Merged

[https://nvbugs/6191524][fix] In MLA.forward_context, also call the warmup when has_cached_kv_for_mla_context #14536

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5e9625b
[nvbugs/6191524][fix] Warm up MLA cat for cached-kv prefill path
tensorrt-cicd May 25, 2026
843bbed
unify code paths and rename functions
zhhuang-nv May 26, 2026
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
7 changes: 3 additions & 4 deletions tensorrt_llm/_torch/attention_backend/trtllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1743,14 +1743,13 @@ def is_chunked_prefill_for_mla_context(
and metadata.num_ctx_cached_tokens > 0
and metadata.runtime_features.chunked_prefill)

def is_chunked_prefill_mla_context_for_warmup(
def has_cached_kv_for_mla_context_warmup(
self,
metadata: TrtllmAttentionMetadata,
) -> bool:
"""Chunked prefill MLA context check for warmup; does not check num_ctx_cached_tokens."""
"""KV cache reuse / chunked prefill MLA context check for warmup, do not check num_ctx_cached_tokens."""
return (self.is_mla_enable and metadata.kv_cache_manager is not None
and metadata.enable_context_mla_with_cached_kv
and metadata.runtime_features.chunked_prefill)
and metadata.enable_context_mla_with_cached_kv)

def load_paged_kv_cache_for_mla(
self,
Expand Down
16 changes: 11 additions & 5 deletions tensorrt_llm/_torch/modules/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2359,9 +2359,11 @@ def forward_context_with_chunked_prefill(

@staticmethod
@functools.cache
def cached_warmup_forward_context_with_chunked_prefill(
num_heads_tp, qk_nope_head_dim, qk_rope_head_dim, kv_lora_rank,
v_head_dim, dtype, device):
def cached_warmup_forward_context_with_cached_kv(num_heads_tp,
qk_nope_head_dim,
qk_rope_head_dim,
kv_lora_rank, v_head_dim,
dtype, device):
"""Warmup torch.compile for cat operations with different tensor layouts.

Tensors are marked with torch._dynamo.maybe_mark_dynamic(..., 0) on the
Expand Down Expand Up @@ -2430,9 +2432,13 @@ def forward_context(
if isinstance(self.mha, TrtllmAttention):
assert isinstance(attn_metadata, TrtllmAttentionMetadata)
trtllm_attention = cast(TrtllmAttention, self.mha)
if trtllm_attention.is_chunked_prefill_mla_context_for_warmup(
# Warm up maybe_compiled_cat for both the chunked-prefill path and
# the cached-kv path; without this, the cached-kv prefill
# (block-reuse without chunked_prefill) recompiles per shape and
# may stall inside inductor's compile worker.
if trtllm_attention.has_cached_kv_for_mla_context_warmup(
attn_metadata):
self.cached_warmup_forward_context_with_chunked_prefill(
self.cached_warmup_forward_context_with_cached_kv(
self.num_heads_tp, self.qk_nope_head_dim,
self.qk_rope_head_dim, self.kv_lora_rank, self.v_head_dim,
q.dtype, q.device)
Expand Down
Loading