From 519cf3ecf148952d7762bd8abed0da609996d9a0 Mon Sep 17 00:00:00 2001 From: amy-why-3459 Date: Thu, 4 Jun 2026 22:49:04 +0800 Subject: [PATCH] [Refactor] [Qwen3-Omni]Modify the thinker's sampling parameters to align with transformers. (#4137) Signed-off-by: amy-why-3459 --- .../perf/tests/test_qwen3_omni_vllm_text.json | 20 ------------------- .../qwen3_omni/run_qwen_omni_acc_benchmark.py | 8 ++++---- tests/test_config_factory.py | 5 ----- vllm_omni/deploy/qwen3_omni_moe.yaml | 6 +----- 4 files changed, 5 insertions(+), 34 deletions(-) diff --git a/tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json b/tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json index 76ac1d858..a973bb0e6 100644 --- a/tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json +++ b/tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json @@ -31,11 +31,6 @@ ], "random_input_len": 2500, "random_output_len": 900, - "temperature": 0.4, - "top_p": 0.9, - "top_k": 1, - "seed": 42, - "repetition_penalty": 1.05, "ignore_eos": true, "extra_body": { "modalities": [ @@ -74,11 +69,6 @@ "random_input_len": 100, "random_output_len": 100, "random_range_ratio": 0.0, - "temperature": 0.4, - "top_p": 0.9, - "top_k": 1, - "seed": 42, - "repetition_penalty": 1.05, "ignore_eos": true, "extra_body": { "modalities": [ @@ -117,11 +107,6 @@ "random_input_len": 100, "random_output_len": 100, "random_range_ratio": 0.0, - "temperature": 0.4, - "top_p": 0.9, - "top_k": 1, - "seed": 42, - "repetition_penalty": 1.05, "ignore_eos": true, "extra_body": { "modalities": [ @@ -162,11 +147,6 @@ "random_input_len": 100, "random_output_len": 100, "random_range_ratio": 0.0, - "temperature": 0.4, - "top_p": 0.9, - "top_k": 1, - "seed": 42, - "repetition_penalty": 1.05, "ignore_eos": true, "extra_body": { "modalities": [ diff --git a/tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py b/tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py index b0dc2af2d..5e872e73a 100644 --- a/tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py +++ b/tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py @@ -8,7 +8,7 @@ L4-style smoke runs against an already-running server. The pytest wrappers in H100 / MI325) because they launch the live Omni server inside the test. 1. **Daily-Omni** — MCQ accuracy fields in the saved JSON (``daily_omni_accuracy``, …); by default the - run **fails** if accuracy is strictly below **0.67** (``--min-daily-omni-accuracy`` / ``ACC_BENCH_MIN_DAILY_OMNI_ACCURACY``). + run **fails** if accuracy is strictly below **0.69** (``--min-daily-omni-accuracy`` / ``ACC_BENCH_MIN_DAILY_OMNI_ACCURACY``). 2. **Seed-TTS** — ``seed-tts-eval``-style metrics when ``--seed-tts-wer-eval`` is used (WER / SIM / UTMOS keys from :func:`compute_seed_tts_wer_metrics`). @@ -308,9 +308,9 @@ def build_arg_parser() -> argparse.ArgumentParser: p.add_argument( "--min-daily-omni-accuracy", type=float, - default=float((os.environ.get("ACC_BENCH_MIN_DAILY_OMNI_ACCURACY") or "0.67").strip() or "0.67"), + default=float((os.environ.get("ACC_BENCH_MIN_DAILY_OMNI_ACCURACY") or "0.69").strip() or "0.69"), help="Fail when daily_omni_accuracy is strictly below this threshold (0–1). " - "Default baseline 0.67; override with env ACC_BENCH_MIN_DAILY_OMNI_ACCURACY or pass 0 to disable the floor.", + "Default baseline 0.69; override with env ACC_BENCH_MIN_DAILY_OMNI_ACCURACY or pass 0 to disable the floor.", ) p.add_argument( @@ -340,7 +340,7 @@ def build_arg_parser() -> argparse.ArgumentParser: p.add_argument( "--max-seed-tts-mean-wer", type=float, - default=0.5, + default=0.25, help="If set, fail when seed_tts_content_error_mean is strictly above this value.", ) p.add_argument( diff --git a/tests/test_config_factory.py b/tests/test_config_factory.py index 9119d78cd..e541d3628 100644 --- a/tests/test_config_factory.py +++ b/tests/test_config_factory.py @@ -1638,9 +1638,6 @@ class TestBaseConfigInheritance: s0 = deploy.stages[0].default_sampling_params # CI overrides max_tokens assert s0["max_tokens"] == 150 - # Inherited from base - assert s0["temperature"] == 0.4 - assert s0["seed"] == 42 def test_pure_inheritance_overlay(self, tmp_path): """An overlay with only ``base_config`` inherits everything.""" @@ -1960,5 +1957,3 @@ class TestSamplingConstraintsPrecedence: assert stages[0].yaml_extras["default_sampling_params"]["detokenize"] is True # Pipeline says stop_token_ids=[2150] for talker assert stages[1].yaml_extras["default_sampling_params"]["stop_token_ids"] == [2150] - # Deploy temperature still flows through - assert stages[0].yaml_extras["default_sampling_params"]["temperature"] == 0.4 diff --git a/vllm_omni/deploy/qwen3_omni_moe.yaml b/vllm_omni/deploy/qwen3_omni_moe.yaml index 3bfa6042a..06869a657 100644 --- a/vllm_omni/deploy/qwen3_omni_moe.yaml +++ b/vllm_omni/deploy/qwen3_omni_moe.yaml @@ -31,12 +31,8 @@ stages: enable_prefix_caching: false devices: "0" default_sampling_params: - temperature: 0.4 - top_p: 0.9 - top_k: 1 + temperature: 0.0 max_tokens: 2048 - seed: 42 - repetition_penalty: 1.05 - stage_id: 1 max_num_batched_tokens: 32768