[Bugfix][Model] Fix DiffusionGemma silently freezing attention mask under CUDA graph replay (#51994)

Signed-off-by: Fabian Joswig <fjosw@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Fabian Joswig
2026-09-30 05:34:26 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 329e6eee3f
commit 1117140edb
3 changed files with 15 additions and 8 deletions
@@ -909,9 +909,12 @@ class DiffusionGemmaModelState(ModelState):
self._req_id_to_index: dict[str, int] = {} self._req_id_to_index: dict[str, int] = {}
# Persistent buffer for per-request causal flags, updated in-place # Persistent buffer for per-request causal flags, updated in-place
# so FULL CUDA graph replay sees the latest values. # so FULL CUDA graph replay sees the latest values. Must be int32:
# FlashAttentionMetadataBuilder.build() rejects other dtypes, since
# an out-of-place cast there would detach the captured graph from
# this buffer and freeze replay at the capture-time snapshot.
self._causal_buf = torch.zeros( self._causal_buf = torch.zeros(
self.max_num_reqs, dtype=torch.bool, device=device self.max_num_reqs, dtype=torch.int32, device=device
) )
# Persistent inputs_embeds buffer — required so FULL CUDA graph # Persistent inputs_embeds buffer — required so FULL CUDA graph
@@ -1116,11 +1119,11 @@ class DiffusionGemmaModelState(ModelState):
# Invariant: the sampler flips is_encoder_phase to False only after a # Invariant: the sampler flips is_encoder_phase to False only after a
# request's FINAL prompt chunk, so a prompt spanning multiple chunks # request's FINAL prompt chunk, so a prompt spanning multiple chunks
# (longer than the token budget) stays causal for every chunk. # (longer than the token budget) stays causal for every chunk.
self._causal_buf[:actual_num_reqs] = self.diffusion_states.is_encoder_phase[ self._causal_buf[:actual_num_reqs].copy_(
slots self.diffusion_states.is_encoder_phase[slots]
] )
if actual_num_reqs < num_reqs: if actual_num_reqs < num_reqs:
self._causal_buf[actual_num_reqs:num_reqs] = False self._causal_buf[actual_num_reqs:num_reqs] = 0
causal: bool | torch.Tensor = self._causal_buf[:num_reqs] causal: bool | torch.Tensor = self._causal_buf[:num_reqs]
return build_attn_metadata( return build_attn_metadata(
+5 -1
View File
@@ -987,7 +987,11 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
scheduler_metadata = self._store_scheduler_metadata(scheduler_metadata) scheduler_metadata = self._store_scheduler_metadata(scheduler_metadata)
if isinstance(causal, torch.Tensor) and causal.dtype != torch.int32: if isinstance(causal, torch.Tensor) and causal.dtype != torch.int32:
causal = causal.to(torch.int32) raise ValueError(
f"Per-request causal tensor must be int32, got {causal.dtype}. "
"Casting here would allocate a fresh tensor each build() and "
"break FULL CUDA graph replay; allocate the buffer as int32."
)
attn_metadata = FlashAttentionMetadata( attn_metadata = FlashAttentionMetadata(
num_actual_tokens=num_actual_tokens, num_actual_tokens=num_actual_tokens,
@@ -217,7 +217,7 @@ def kernel_unified_attention(
SLIDING_WINDOW: tl.constexpr, # int SLIDING_WINDOW: tl.constexpr, # int
USE_CAUSAL: tl.constexpr, # bool USE_CAUSAL: tl.constexpr, # bool
USE_PER_SEQ_CAUSAL: tl.constexpr, # bool USE_PER_SEQ_CAUSAL: tl.constexpr, # bool
per_seq_causal_ptr, # [num_seqs] bool, or None per_seq_causal_ptr, # [num_seqs] int32, or None
USE_MM_PREFIX: tl.constexpr, # bool USE_MM_PREFIX: tl.constexpr, # bool
MAX_MM_RANGES: tl.constexpr, # int MAX_MM_RANGES: tl.constexpr, # int
mm_prefix_range_ptr, mm_prefix_range_ptr,