[Perf] Add fused DFlash2 grouped convolution (#55960)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Misha Goin
2026-09-17 21:38:59 +00:00
committed by GitHub
parent 668d6c3a77
commit db7a24c230
4 changed files with 199 additions and 0 deletions
@@ -26,6 +26,8 @@ class DFlashCorrectnessConfig:
use_chat_completions: bool = False
enforce_eager: bool = False
disable_flashinfer_sampler: bool = False
language_model_only: bool = False
chat_template_kwargs: dict[str, object] | None = None
QWEN3_DFLASH = DFlashCorrectnessConfig(
@@ -35,6 +37,18 @@ QWEN3_DFLASH = DFlashCorrectnessConfig(
expected_acceptance_len=3.5,
)
QWEN3_8_DFLASH2_NVFP4 = DFlashCorrectnessConfig(
model="nvidia/Qwen3.8-27B-NVFP4",
draft_model="z-lab/Qwen3.8-27B-DFlash2",
expected_accuracy=0.90,
expected_acceptance_len=6.14 * 0.95,
num_speculative_tokens=7,
max_num_seqs=32,
use_chat_completions=True,
language_model_only=True,
chat_template_kwargs={"enable_thinking": False},
)
LAGUNA_DFLASH_NVFP4 = DFlashCorrectnessConfig(
model="poolside/Laguna-XS-2.1-NVFP4",
draft_model="poolside/Laguna-XS-2.1-DFlash-NVFP4",
@@ -104,6 +118,11 @@ def test_dflash_reference_acceptance_lengths(
True,
id="laguna-nvfp4-mrv2",
),
pytest.param(
QWEN3_8_DFLASH2_NVFP4,
True,
id="qwen3.8-dflash2-nvfp4-mrv2",
),
],
)
def test_dflash_correctness(
@@ -134,12 +153,14 @@ def test_dflash_correctness(
disable_log_stats=False,
enable_chunked_prefill=None,
compilation_config=CompilationConfig(),
language_model_only=config.language_model_only,
) as spec_runner:
spec_llm = spec_runner.llm
results = evaluate_gsm8k_offline(
spec_llm,
num_questions=config.num_questions,
use_chat_completions=config.use_chat_completions,
chat_template_kwargs=config.chat_template_kwargs,
)
accuracy = results["accuracy"]
acceptance_len = compute_acceptance_len(spec_llm.get_metrics())