Skip to content

[None][fix] Use 128-token DSpark draft pages only where the dynamic-page kernels are verified (SM100/SM103) #23

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

Closed
zheyuf wants to merge 1 commit into from
Closed

[None][fix] Use 128-token DSpark draft pages only where the dynamic-page kernels are verified (SM100/SM103) #23

Show file tree
Hide file tree
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
5 changes: 4 additions & 1 deletion tensorrt_llm/_torch/speculative/dflash_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,6 +168,8 @@ def dflash_trtllm_supports_page_size(tokens_per_block: int, *, has_context_atten
not has_context_attention
and is_power_of_two
and tokens_per_block >= _TRTLLM_GEN_DYNAMIC_PAGE_MIN_TOKENS
# The dynamic page-size kernels are verified on SM100 and SM103 only.
and get_sm_version() in (100, 103)
)


Expand Down Expand Up @@ -206,7 +208,8 @@ def validate_dflash_trtllm_gen_runtime(
raise RuntimeError(
"DFlash TRTLLM attention does not support context-cache page size "
f"{tokens_per_block}; supported sizes are {supported}, plus powers of two "
f">= {_TRTLLM_GEN_DYNAMIC_PAGE_MIN_TOKENS} without the non-causal context kernel."
f">= {_TRTLLM_GEN_DYNAMIC_PAGE_MIN_TOKENS} on SM100/SM103 without the non-causal "
"context kernel."
)
if has_context_attention and head_dim in FlashInferTrtllmGenFmha.UNSUPPORTED_HEAD_SIZES_CONTEXT:
raise RuntimeError(
Expand Down
24 changes: 18 additions & 6 deletions tests/unittest/_torch/executor/kv_cache/test_kv_cache_estimation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1514,17 +1514,29 @@ def test_premise_trtllm_gen_page_sizes(self):
assert max(size for size in supported if size < 128) == 64

@pytest.mark.parametrize(
"tokens_per_block, has_context_attention, expected",
[(64, True, True), (128, False, True), (128, True, False), (96, False, False)],
"tokens_per_block, has_context_attention, sm_version, expected",
[
(64, True, 103, True),
(128, False, 103, True),
(128, False, 107, False),
(128, True, 103, False),
(96, False, 103, False),
],
)
def test_trtllm_page_size_capability(self, tokens_per_block, has_context_attention, expected):
def test_trtllm_page_size_capability(
self, tokens_per_block, has_context_attention, sm_version, expected
):
from tensorrt_llm._torch.speculative.dflash_attention import (
dflash_trtllm_supports_page_size,
)

supported = dflash_trtllm_supports_page_size(
tokens_per_block, has_context_attention=has_context_attention
)
with patch(
"tensorrt_llm._torch.speculative.dflash_attention.get_sm_version",
return_value=sm_version,
):
supported = dflash_trtllm_supports_page_size(
tokens_per_block, has_context_attention=has_context_attention
)
assert supported is expected

@pytest.mark.parametrize(
Expand Down
Loading