[Spec Decode] Register Qwen3.6 dSpark acceptance coverage (#51310)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Michael Goin
2026-08-07 17:07:55 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 9823714346
commit c8fe1d5715
2 changed files with 122 additions and 100 deletions
+7 -1
View File
@@ -292,6 +292,7 @@ def evaluate_gsm8k_offline(
temperature: float = 0.0,
gen_prefix: str = "",
use_chat_completions: bool = False,
chat_template_kwargs: dict[str, object] | None = None,
) -> dict[str, float | int]:
"""Evaluate GSM8K accuracy using an offline vllm.LLM object.
@@ -300,6 +301,7 @@ def evaluate_gsm8k_offline(
When ``use_chat_completions=True``, prompts go through the chat template via
``llm.chat()`` instead of raw completion (for instruction-tuned models).
``chat_template_kwargs`` are forwarded to ``llm.chat()`` when provided.
"""
from vllm import SamplingParams
@@ -318,7 +320,11 @@ def evaluate_gsm8k_offline(
tic = time.perf_counter()
if use_chat_completions:
conversations = [[{"role": "user", "content": p}] for p in prompts]
outputs = llm.chat(conversations, sampling_params)
outputs = llm.chat(
conversations,
sampling_params,
chat_template_kwargs=chat_template_kwargs,
)
else:
outputs = llm.generate(prompts, sampling_params)
latency = time.perf_counter() - tic
@@ -1,6 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
import pytest
import torch
@@ -10,120 +12,134 @@ from vllm.distributed import cleanup_dist_env_and_memory
from ...utils import compute_acceptance_len, compute_acceptance_rate
REGRESSION_TOLERANCE = 0.95
def test_gemma4_dspark_correctness_and_acceptance_rate(
@dataclass(frozen=True)
class DSparkCorrectnessConfig:
model: str
draft_model: str
reference_accuracy: float
reference_acceptance_rate: float
reference_acceptance_len: float
num_speculative_tokens: int = 7
max_model_len: int = 4096
max_num_seqs: int | None = None
num_questions: int = 1319
max_tokens: int = 256
use_chat_completions: bool = False
chat_template_kwargs: dict[str, object] | None = None
attention_backend: str | None = None
gpu_memory_utilization: float = 0.92
enforce_eager: bool = False
disable_flashinfer_sampler: bool = False
language_model_only: bool = False
# References from 12 full GSM8K runs at temperature 1.0: accuracy 0.782-0.814,
# acceptance rate 0.418-0.434, acceptance length 3.928-4.037.
QWEN3_DSPARK_DEEPSPEC = DSparkCorrectnessConfig(
model="Qwen/Qwen3-4B-FP8",
draft_model="deepseek-ai/dspark_qwen3_4b_block7",
reference_accuracy=0.801,
reference_acceptance_rate=0.428,
reference_acceptance_len=3.994,
)
# References from five 200-question runs at temperature 1.0: accuracy
# 0.900-0.955, acceptance rate 0.578-0.595, acceptance length 5.044-5.167.
GEMMA4_DSPARK_DEEPSPEC = DSparkCorrectnessConfig(
model="RedHatAI/gemma-4-12B-it-NVFP4",
draft_model="deepseek-ai/dspark_gemma4_12b_block7",
reference_accuracy=0.92,
reference_acceptance_rate=0.58,
reference_acceptance_len=5.0,
max_num_seqs=32,
num_questions=200,
use_chat_completions=True,
gpu_memory_utilization=0.85,
)
# Reference from 200 GSM8K questions at temperature 1.0.
QWEN3_6_DSPARK_SPECULATORS = DSparkCorrectnessConfig(
model="RedHatAI/Qwen3.6-35B-A3B-NVFP4",
draft_model="RedHatAI/Qwen3.6-35B-A3B-speculator.dspark",
reference_accuracy=0.85,
reference_acceptance_rate=0.41,
reference_acceptance_len=4.25,
num_speculative_tokens=8,
max_num_seqs=32,
num_questions=200,
use_chat_completions=True,
chat_template_kwargs={"enable_thinking": False},
gpu_memory_utilization=0.85,
language_model_only=True,
)
@pytest.mark.parametrize(
"config",
[
pytest.param(QWEN3_DSPARK_DEEPSPEC, id="qwen3-deepspec"),
pytest.param(GEMMA4_DSPARK_DEEPSPEC, id="gemma4-deepspec"),
pytest.param(QWEN3_6_DSPARK_SPECULATORS, id="qwen3.6-speculators"),
],
)
def test_dspark_correctness_and_acceptance_rate(
monkeypatch: pytest.MonkeyPatch,
config: DSparkCorrectnessConfig,
):
"""
E2E test for Gemma4 DSpark speculative decoding: acceptance rate/length
regression coverage plus GSM8K correctness, at temperature=1.0 to exercise
the probabilistic draft-sampling/rejection-sampling path (not just greedy).
"""Guard GSM8K accuracy and acceptance metrics for DSpark models."""
speculative_config = {
"method": "dspark",
"model": config.draft_model,
"num_speculative_tokens": config.num_speculative_tokens,
"draft_sample_method": "probabilistic",
}
if config.attention_backend is not None:
speculative_config["attention_backend"] = config.attention_backend
gemma-4-12B is instruct-tuned, so GSM8K is run through the chat template
(use_chat_completions=True; raw few-shot completion collapses to a few
percent). Reference: measured over 5 runs of 200 GSM8K questions at
temperature=1.0 (prefix caching disabled):
accuracy: min=0.900 max=0.955 mean=0.937
acceptance_rate: min=0.578 max=0.595 mean=0.588
acceptance_len: min=5.044 max=5.167 mean=5.116
Thresholds set conservatively to 10% to avoid flaking due to unlucky sampling
"""
monkeypatch.setenv("VLLM_USE_FLASHINFER_SAMPLER", "0")
llm_kwargs = {
"model": config.model,
"trust_remote_code": True,
"speculative_config": speculative_config,
"max_model_len": config.max_model_len,
"gpu_memory_utilization": config.gpu_memory_utilization,
"enforce_eager": config.enforce_eager,
"enable_prefix_caching": False,
"disable_log_stats": False,
}
if config.max_num_seqs is not None:
llm_kwargs["max_num_seqs"] = config.max_num_seqs
if config.language_model_only:
llm_kwargs["language_model_only"] = True
spec_llm = LLM(
model="RedHatAI/gemma-4-12B-it-NVFP4",
trust_remote_code=True,
speculative_config={
"method": "dspark",
"model": "deepseek-ai/dspark_gemma4_12b_block7",
"num_speculative_tokens": 7,
"draft_sample_method": "probabilistic",
},
max_model_len=4096,
max_num_seqs=32,
gpu_memory_utilization=0.85,
enforce_eager=True,
enable_prefix_caching=False,
disable_log_stats=False,
)
spec_llm = LLM(**llm_kwargs)
try:
results = evaluate_gsm8k_offline(
spec_llm, num_questions=200, temperature=1.0, use_chat_completions=True
spec_llm,
num_questions=config.num_questions,
max_tokens=config.max_tokens,
temperature=1.0,
use_chat_completions=config.use_chat_completions,
chat_template_kwargs=config.chat_template_kwargs,
)
gsm8k_accuracy = results["accuracy"]
accuracy = results["accuracy"]
metrics = spec_llm.get_metrics()
acceptance_rate = compute_acceptance_rate(metrics)
acceptance_len = compute_acceptance_len(metrics)
print(
f"Gemma4 DSpark acceptance_rate={acceptance_rate:.2f}, "
f"acceptance_len={acceptance_len:.2f}, "
f"gsm8k_accuracy={gsm8k_accuracy:.3f}"
f"{config.model}: GSM8K accuracy={accuracy:.3f}, "
f"acceptance_rate={acceptance_rate:.2f}, "
f"acceptance_len={acceptance_len:.2f}"
)
assert acceptance_rate >= 0.58 * 0.9
assert acceptance_len >= 5.0 * 0.9
assert gsm8k_accuracy >= 0.92 * 0.9
assert accuracy >= config.reference_accuracy * REGRESSION_TOLERANCE
assert (
acceptance_rate >= config.reference_acceptance_rate * REGRESSION_TOLERANCE
)
assert acceptance_len >= config.reference_acceptance_len * REGRESSION_TOLERANCE
finally:
del spec_llm
torch.accelerator.empty_cache()
cleanup_dist_env_and_memory()
@pytest.fixture
def dspark_config():
target_model = "Qwen/Qwen3-4B-FP8"
draft_model = "deepseek-ai/dspark_qwen3_4b_block7"
return dict(
model=target_model,
trust_remote_code=True,
speculative_config={
"method": "dspark",
"model": draft_model,
"num_speculative_tokens": 7,
"attention_backend": "FLASH_ATTN",
"draft_sample_method": "probabilistic",
},
max_model_len=4096,
disable_log_stats=False,
)
def test_dspark_correctness_and_acceptance_rate(dspark_config):
"""
E2E test for DSpark speculative decoding: acceptance rate/length
regression coverage plus GSM8K correctness, at temperature=1.0 to
exercise the probabilistic draft-sampling/rejection-sampling path
(not just greedy).
Uses Qwen/Qwen3-4B-FP8 as target with the dspark_qwen3_4b_block7 draft
model. Reference: measured over 12 runs of the full GSM8K set at
temperature=1.0 (prefix caching disabled to avoid cross-run reuse):
accuracy: min=0.782 max=0.814 mean=0.801
acceptance_rate: min=0.418 max=0.434 mean=0.428
acceptance_len: min=3.928 max=4.037 mean=3.994
Thresholds set conservatively to 10% to avoid flaking due to unlucky sampling
"""
spec_llm = LLM(**dspark_config)
results = evaluate_gsm8k_offline(spec_llm, temperature=1.0)
gsm8k_accuracy = results["accuracy"]
metrics = spec_llm.get_metrics()
acceptance_rate = compute_acceptance_rate(metrics)
acceptance_len = compute_acceptance_len(metrics)
print(
f"DSpark acceptance_rate={acceptance_rate:.2f}, "
f"acceptance_len={acceptance_len:.2f}, "
f"gsm8k_accuracy={gsm8k_accuracy:.3f}"
)
assert acceptance_rate >= 0.428 * 0.9
assert acceptance_len >= 3.994 * 0.9
assert gsm8k_accuracy >= 0.801 * 0.9
del spec_llm
torch.accelerator.empty_cache()
cleanup_dist_env_and_memory()