[Bugfix] Honor STEP token pooling in DispatchPooler.for_seq_cls (#55307)

Signed-off-by: weihan <130119942+krsish@users.noreply.github.com>
This commit is contained in:
weihan
2026-09-08 04:13:09 +00:00
committed by GitHub
parent 7fa2c63796
commit 9ea8f3ffc3
3 changed files with 73 additions and 47 deletions
@@ -4,20 +4,25 @@
# https://github.com/QwenLM/Qwen3-ASR
"""
Online forced alignment example using Qwen3-ForcedAligner-0.6B.
Online forced alignment example using Qwen3-ForcedAligner-0.6B-hf.
Forced alignment takes audio and reference text as input and produces
word-level timestamps. The model predicts a time bin at each <timestamp>
token position; multiplying by ``timestamp_segment_time`` gives milliseconds.
This example uses STEP pooling to return logits only for <timestamp> tokens.
The ``step_tag_id`` value 151705 is this checkpoint's ``timestamp_token_id``
from its config.json.
Start the server with:
vllm serve Qwen/Qwen3-ForcedAligner-0.6B \\
--runner pooling \\
--enforce-eager \\
--trust-request-chat-template \\
--hf-overrides \\
'{"architectures": ["Qwen3ASRForcedAlignerForTokenClassification"]}'
vllm serve Qwen/Qwen3-ForcedAligner-0.6B-hf \
--runner pooling \
--convert classify \
--pooler-config \
'{"tok_pooling_type":"STEP","step_tag_id":151705,"use_activation":false}' \
--enforce-eager \
--trust-request-chat-template
Then run:
@@ -108,7 +113,7 @@ def parse_response(response: requests.Response) -> dict[str, Any]:
return result
def load_timestamp_config(model: str) -> tuple[int, float]:
def load_timestamp_segment_time(model: str) -> float:
model_path = Path(model)
config_path = (
model_path / "config.json"
@@ -119,7 +124,7 @@ def load_timestamp_config(model: str) -> tuple[int, float]:
with config_path.open() as f:
config = json.load(f)
return config["timestamp_token_id"], config["timestamp_segment_time"]
return config["timestamp_segment_time"]
def parse_args():
@@ -129,7 +134,7 @@ def parse_args():
parser.add_argument(
"--model",
type=str,
default="Qwen/Qwen3-ForcedAligner-0.6B",
default="Qwen/Qwen3-ForcedAligner-0.6B-hf",
)
parser.add_argument(
"--audio-path",
@@ -147,8 +152,6 @@ def parse_args():
def main(args):
from transformers import AutoTokenizer
api_url = f"http://{args.host}:{args.port}/pooling"
prompt = build_prompt(args.words)
audio_uri = (
@@ -161,46 +164,22 @@ def main(args):
pooling_response = post_http_request(payload=payload, api_url=api_url)
result = parse_response(pooling_response)
tokenizer = AutoTokenizer.from_pretrained(args.model)
timestamp_token_id, timestamp_segment_time = load_timestamp_config(args.model)
timestamp_segment_time = load_timestamp_segment_time(args.model)
output = result["data"][0]
logits = torch.tensor(output["data"])
predictions = logits.argmax(dim=-1)
token_ids = tokenizer(prompt, add_special_tokens=False)["input_ids"]
audio_pad_token_id = tokenizer.convert_tokens_to_ids("<|audio_pad|>")
usage = result.get("usage") or {}
prompt_tokens = usage.get("prompt_tokens")
if prompt_tokens is not None and prompt_tokens != len(predictions):
expected_timestamps = len(args.words) * 2
if len(predictions) != expected_timestamps:
raise RuntimeError(
"The response length does not match the reported prompt token count."
f"Expected {expected_timestamps} timestamp predictions, "
f"but received {len(predictions)}. Check that the server was started "
"with STEP pooling."
)
try:
audio_pad_index = token_ids.index(audio_pad_token_id)
except ValueError as exc:
raise RuntimeError("The prompt does not contain the audio pad token.") from exc
audio_token_shift = len(predictions) - len(token_ids)
if audio_token_shift < 0:
raise RuntimeError(
"The response is shorter than the locally tokenized prompt. "
"Check that the server was started with --trust-request-chat-template."
)
ts_predictions = []
for i, token_id in enumerate(token_ids):
if token_id != timestamp_token_id:
continue
prediction_index = i + audio_token_shift if i > audio_pad_index else i
ts_predictions.append(
predictions[prediction_index].item() * timestamp_segment_time
)
if len(ts_predictions) < len(args.words) * 2:
raise RuntimeError("The model did not return enough timestamp predictions.")
ts_predictions = [
prediction.item() * timestamp_segment_time for prediction in predictions
]
for i, word in enumerate(args.words):
start_ms = ts_predictions[i * 2]