[MRV2][Spec Decode] DFlash (#44586)

Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai>
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Signed-off-by: Benjamin Chislett <chislett.ben@gmail.com>
Co-authored-by: Giancarlo Delfin <gdelfin@inferact.ai>
This commit is contained in:
Benjamin Chislett
2026-06-10 08:47:46 -07:00
committed by GitHub
co-authored by Giancarlo Delfin
parent 4673ca1d78
commit 0bae1d3848
14 changed files with 868 additions and 47 deletions
+12 -2
View File
@@ -1300,13 +1300,18 @@ def dflash_config():
)
def test_dflash_acceptance_rates(dflash_config):
@pytest.mark.parametrize("use_mrv2", [False, True])
def test_dflash_acceptance_rates(
monkeypatch: pytest.MonkeyPatch, use_mrv2: bool, dflash_config
):
"""
E2E test for DFlash (block diffusion) speculative decoding.
Runs acceptance rate validation on GSM8k, MT-Bench, and HumanEval
comparing against baseline results from the paper (Table 1).
See https://github.com/z-lab/dflash/blob/main/benchmark_sglang.py for methodology.
"""
monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if use_mrv2 else "0")
spec_llm = LLM(**dflash_config)
max_prompts_per_dataset = 200 # mt-bench has 80, humaneval has 164, truncates gsm8k
@@ -1414,11 +1419,16 @@ def test_synthetic_acceptance_rate():
cleanup_dist_env_and_memory()
def test_dflash_correctness(dflash_config):
@pytest.mark.parametrize("use_mrv2", [False, True])
def test_dflash_correctness(
monkeypatch: pytest.MonkeyPatch, use_mrv2: bool, dflash_config
):
"""
E2E test for DFlash (block diffusion) speculative decoding.
Ensures output correctness on GSM8k, with cudagraphs and batching on.
"""
monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if use_mrv2 else "0")
spec_llm = LLM(**dflash_config)
# Evaluate GSM8k accuracy (Qwen3-8B ref: ~87-92% on GSM8k)