[CI/Build][MiniCPM-o] Fix the Nightly tests (#7758)

Signed-off-by: chickeyton <ngton2014@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
chickeytonandClaude Fable 5.1 authored and GitHub committed 2026-09-18 08:58:33 +00:00
1 parent aa2fadac42
commit 6df62d5157
8 files changed
+233 -18

No files matched your search

+81 -5
View File
@@ -26,8 +26,8 @@
{
"dataset_name": "seed-tts",
"dataset_path": "zhaochenyang20/seed-tts-eval",
"backend": "openai-realtime-tts",
"endpoint": "/v1/realtime",
"backend": "openai-chat-omni",
"endpoint": "/v1/chat/completions",
"num_prompts": [
32,
64,
@@ -42,7 +42,73 @@
"disable_shuffle": true,
"trust_remote_code": true,
"seed_tts_locale": "en",
"percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration"
"extra_body": {
"modalities": [
"text",
"audio"
],
"chat_template_kwargs": {
"enable_thinking": false,
"use_tts_template": true
}
},
"percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
"baseline": {
"H100": {
"request_throughput": [
0.9731,
1.1986,
1.686
],
"mean_ttft_ms": [
152.5665,
270.2559,
301.455
],
"mean_e2el_ms": [
1027.2953,
3306.3735,
4686.5366
],
"mean_audio_ttfp_ms": [
458.1553,
2030.2335,
2068.6233
],
"mean_audio_rtf": [
0.2357,
0.791,
1.0324
]
},
"A3": {
"request_throughput": [
0.5383,
0.6042,
0.7547
],
"mean_ttft_ms": [
333.2633,
533.8656,
547.3388
],
"mean_e2el_ms": [
1857.2154,
6600.5101,
10450.6811
],
"mean_audio_ttfp_ms": [
986.4666,
3411.0754,
3352.7515
],
"mean_audio_rtf": [
0.4423,
1.5734,
2.3024
]
}
}
}
]
},
@@ -72,8 +138,8 @@
{
"dataset_name": "seed-tts",
"dataset_path": "zhaochenyang20/seed-tts-eval",
"backend": "openai-realtime-tts",
"endpoint": "/v1/realtime",
"backend": "openai-chat-omni",
"endpoint": "/v1/chat/completions",
"num_prompts": [
32,
64,
@@ -88,6 +154,16 @@
"disable_shuffle": true,
"trust_remote_code": true,
"seed_tts_locale": "en",
"extra_body": {
"modalities": [
"text",
"audio"
],
"chat_template_kwargs": {
"enable_thinking": false,
"use_tts_template": true
}
},
"percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration"
}
]
@@ -26,8 +26,8 @@
{
"dataset_name": "seed-tts",
"dataset_path": "zhaochenyang20/seed-tts-eval",
"backend": "openai-realtime-tts",
"endpoint": "/v1/realtime",
"backend": "openai-chat-omni",
"endpoint": "/v1/chat/completions",
"num_prompts": [
10,
40
@@ -40,7 +40,41 @@
"disable_shuffle": true,
"trust_remote_code": true,
"seed_tts_locale": "en",
"percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration"
"extra_body": {
"modalities": [
"text",
"audio"
],
"chat_template_kwargs": {
"enable_thinking": false,
"use_tts_template": true
}
},
"percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
"baseline": {
"H100": {
"request_throughput": [
0.9731,
1.1986
],
"mean_ttft_ms": [
152.5665,
270.2559
],
"mean_e2el_ms": [
1027.2953,
3306.3735
],
"mean_audio_ttfp_ms": [
458.1553,
2030.2335
],
"mean_audio_rtf": [
0.2357,
0.791
]
}
}
}
]
}
@@ -29,7 +29,16 @@ _SERVER_ENV: dict[str, str] = {"VLLM_OMNI_INPUT_WAIT_TIMEOUT_S": str(SERVER_INPU
if TOKENIZER:
_SERVER_ENV["NEMOTRON_VOICECHAT_LLM_PATH"] = TOKENIZER
pytestmark = [pytest.mark.full_model, pytest.mark.omni]
pytestmark = [
pytest.mark.full_model,
pytest.mark.omni,
# Since #7413 a duplex-capable model is served by the duplex engine only when
# its pipeline declares a duplex_plugin. The Nemotron VoiceChat pipeline does
# not, so the server starts the turn engine and the /v1/realtime websocket
# falls through to vllm's speech-to-text realtime handler, which rejects the
# native duplex client's session.update ("Missing required field: model").
pytest.mark.skip(reason="Nemotron VoiceChat is not served by the duplex engine yet (no duplex_plugin)"),
]
@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
@@ -20,12 +20,14 @@ from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_llm import (
from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts import (
_CODEC_PENALTY_WINDOW,
_DUPLEX_CODEC_TOKENS_PER_CHUNK,
_DUPLEX_TURN_END_CODEC_TOKENS,
_OFFLINE_CODEC_MAX_NEW_TOKENS,
_REPETITION_PENALTY_CHUNK_SIZE,
MiniCPMO45OmniTTSForConditionalGeneration,
_apply_batched_repetition_penalty,
_native_duplex_chunk_budget,
_restore_weight_norm_weight,
_turn_end_boundary_eos_masked,
blank_scheduler_prompt_for_penalties,
)
from vllm_omni.utils.mm_outputs import to_payload_element
@@ -453,7 +455,9 @@ def test_sample_leaves_unforced_rows_to_the_sampler(mocker) -> None:
def test_native_duplex_chunk_budget_masks_eos_until_generate_chunk_is_full() -> None:
assert _native_duplex_chunk_budget({}) == (_DUPLEX_CODEC_TOKENS_PER_CHUNK, _DUPLEX_CODEC_TOKENS_PER_CHUNK)
assert _native_duplex_chunk_budget({"turn_start": True}) == (_DUPLEX_CODEC_TOKENS_PER_CHUNK, 0)
assert _native_duplex_chunk_budget({"turn_end": True}) == (_DUPLEX_CODEC_TOKENS_PER_CHUNK, 0)
# The turn-end chunk drains the owed text: no floor, multi-unit ceiling.
assert _native_duplex_chunk_budget({"turn_end": True}) == (_DUPLEX_TURN_END_CODEC_TOKENS, 0)
assert _DUPLEX_TURN_END_CODEC_TOKENS > _DUPLEX_CODEC_TOKENS_PER_CHUNK
talker = _make_talker()
talker._request_audio_states["req-chunk"] = {
@@ -1097,3 +1101,43 @@ def test_request_cleanup_evicts_decode_state() -> None:
assert "req-done" not in talker._request_audio_states
assert "req-done" not in talker._request_condition_states
def test_turn_end_drain_masks_only_the_cadence_eos() -> None:
# Early EOS (text spoken before the first boundary) is allowed.
assert not _turn_end_boundary_eos_masked(0)
assert not _turn_end_boundary_eos_masked(18)
assert not _turn_end_boundary_eos_masked(24)
# The cadence EOS right after each 25-frame boundary is masked for a short window.
assert _turn_end_boundary_eos_masked(25)
assert _turn_end_boundary_eos_masked(29)
assert not _turn_end_boundary_eos_masked(30)
assert not _turn_end_boundary_eos_masked(49)
assert _turn_end_boundary_eos_masked(50)
assert not _turn_end_boundary_eos_masked(60)
talker = _make_talker()
talker._request_audio_states["req-drain"] = {
"finished": False,
"step": 25,
"max_tokens": _DUPLEX_TURN_END_CODEC_TOKENS,
"min_tokens": 0,
"turn_end_drain": True,
}
output = talker.make_omni_output(
torch.ones(1, 2),
model_intermediate_buffer=[{"request_id": "req-drain", "codes": {"audio": torch.empty(0)}}],
request_token_spans=[(0, 1)],
)
logits = talker.compute_logits(output.text_hidden_states)
assert logits[0, 7].item() == float("-inf")
assert output.multimodal_outputs["meta"]["finished"][0].item() is False
talker._request_audio_states["req-drain"]["step"] = 30
output = talker.make_omni_output(
torch.ones(1, 2),
model_intermediate_buffer=[{"request_id": "req-drain", "codes": {"audio": torch.empty(0)}}],
request_token_spans=[(0, 1)],
)
logits = talker.compute_logits(output.text_hidden_states)
assert torch.isfinite(logits[0, 7])