anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity

[Refactor] Output Processor Phase 2: separate multimodal output channel (#1601) (#2744)

Browse Source 
Signed-off-by: Megha Agarwal <agarwalmegha1308@gmail.com>
Signed-off-by: meghaagr13 <divyanshsinghvi@gmail.com>
Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Co-authored-by: meghaagr13 <divyanshsinghvi@gmail.com>
Co-authored-by: Boao Shi <aoibosh@connect.hku.hk>
Co-authored-by: Taichang Zhou <tzhouam@connect.ust.hk>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Megha AgarwalCopilotHongsheng Liumeghaagr13Boao ShiTaichang ZhouCursor
 authored and GitHub committed 2026-06-12 15:19:14 +08:00
1 parent 1be1a71156
commit 8f869c8ea5
49 files changed
+985 -327
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

tests/model_executor/stage_input_processors/test_cosyvoice3_stage_input_processors.py
+12 -12
View File
@@ -140,7 +140,7 @@ def test_talker2code2wav_async_chunk_final_payload_uses_absolute_token_offset():
payload = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=True,
)
@@ -167,7 +167,7 @@ def test_talker2code2wav_async_chunk_emits_eof_when_finished_without_valid_codes
payload = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=True,
)
@@ -188,13 +188,13 @@ def test_talker2code2wav_async_chunk_does_not_reemit_without_new_tokens():
payload1 = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
payload2 = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
@@ -216,14 +216,14 @@ def test_talker2code2wav_async_chunk_waits_for_prelookahead_and_emits_cumulative
payload_pending = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
request.output_token_ids = [1, 2, 3]
payload_ready = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
@@ -246,14 +246,14 @@ def test_talker2code2wav_async_chunk_final_flush_uses_previous_token_offset():
payload_stream = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
request.output_token_ids = [3, 4, 5, 6]
payload_final = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=True,
)
@@ -281,14 +281,14 @@ def test_talker2code2wav_async_chunk_respects_prompt_token_pad_on_first_chunk():
payload_pending = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
request.output_token_ids = [8, 9, 10, 11]
payload_ready = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
@@ -310,13 +310,13 @@ def test_talker2code2wav_async_chunk_emits_terminal_eof_without_duplicate_audio(
payload_stream = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=False,
)
payload_final = talker2code2wav_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
is_finished=True,
)
tests/model_executor/stage_input_processors/test_qwen3_tts_async_chunk.py
+10 -10
View File
@@ -56,7 +56,7 @@ def _call(tm, rid, *, n_frames, finished=False, req_ic=None):
tm.code_prompt_token_ids[rid] = [_FRAME[:] for _ in range(n_frames)]
return talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.zeros((0,))}},
multimodal_output={"codes": {"audio": torch.zeros((0,))}},
request=_req(rid, finished=finished, initial_codec_chunk_frames=req_ic),
is_finished=finished,
)
@@ -66,7 +66,7 @@ def test_empty_returns_none():
tm = _tm()
p = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.zeros((0,))}},
multimodal_output={"codes": {"audio": torch.zeros((0,))}},
request=_req("r", finished=False),
)
assert p is None
@@ -76,7 +76,7 @@ def test_eof_marker_when_finished_empty():
tm = _tm()
p = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output=None,
multimodal_output=None,
request=_req("r", finished=True),
is_finished=True,
)
@@ -89,7 +89,7 @@ def test_flush_on_finish():
tm.code_prompt_token_ids["r"] = [_FRAME[:] for _ in range(24)]
p = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output=None,
multimodal_output=None,
request=_req("r", finished=True),
is_finished=True,
)
@@ -262,7 +262,7 @@ def test_first_streaming_chunk_prepends_ref_code_context():
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
multimodal_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
@@ -286,7 +286,7 @@ def test_followup_ref_code_context_is_sent_as_metadata_handle():
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
multimodal_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
@@ -319,7 +319,7 @@ def test_streaming_ref_code_context_is_bounded_for_batchable_shapes():
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
multimodal_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
request=_req(rid, finished=False),
is_finished=False,
)
@@ -338,7 +338,7 @@ def test_ref_code_context_can_be_buffered_before_first_emit():
first_payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]]), "ref": ref_code}},
multimodal_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]]), "ref": ref_code}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
@@ -348,14 +348,14 @@ def test_ref_code_context_can_be_buffered_before_first_emit():
for _ in range(8):
talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]])}},
multimodal_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]])}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]])}},
multimodal_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]])}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
tests/model_executor/stage_input_processors/test_voxtral_tts_async_chunk.py
+14 -14
View File
@@ -92,7 +92,7 @@ def test_empty_chunk_when_not_finished():
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output={"audio": torch.zeros((0,))},
multimodal_output={"audio": torch.zeros((0,))},
request=request,
)
@@ -109,7 +109,7 @@ def test_flush_tail_when_finished():
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None, # e.g. EOS step with no audio
multimodal_output=None, # e.g. EOS step with no audio
request=request,
)
@@ -133,7 +133,7 @@ def test_eof_marker_when_finished_with_no_frames():
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
)
@@ -153,10 +153,10 @@ def test_normal_chunk_emission():
# Feed 25 frames one by one
for i in range(25):
pooling_output = {"audio": torch.tensor([float(i)] * 4)}
multimodal_output = {"audio": torch.tensor([float(i)] * 4)}
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=pooling_output,
multimodal_output=multimodal_output,
request=request,
)
@@ -181,10 +181,10 @@ def test_small_initial_chunks():
# Feed 5 frames (should trigger emission because codec_chunk_frames_at_begin=5)
for i in range(5):
pooling_output = {"audio": torch.tensor([float(i + 1)] * 3)}
multimodal_output = {"audio": torch.tensor([float(i + 1)] * 3)}
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=pooling_output,
multimodal_output=multimodal_output,
request=request,
)
@@ -208,10 +208,10 @@ def test_no_emission_between_boundaries():
# Feed 3 frames (3 % 5 != 0, should not emit)
for i in range(3):
pooling_output = {"audio": torch.tensor([float(i)] * 4)}
multimodal_output = {"audio": torch.tensor([float(i)] * 4)}
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=pooling_output,
multimodal_output=multimodal_output,
request=request,
)
@@ -230,10 +230,10 @@ def test_context_handling_format():
# Feed 5 frames to trigger a chunk (chunk_size_at_begin=5)
for i in range(5):
pooling_output = {"audio": torch.tensor([float(i + 10)] * 2)} # codebook_dim=2
multimodal_output = {"audio": torch.tensor([float(i + 10)] * 2)} # codebook_dim=2
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=pooling_output,
multimodal_output=multimodal_output,
request=request,
)
@@ -250,14 +250,14 @@ def test_context_handling_format():
assert flat_codes.numel() == total_window_frames * 2 # codebook_dim=2
def test_none_pooling_output_not_finished_returns_none():
"""None pooling_output when not finished returns None."""
def test_none_multimodal_output_not_finished_returns_none():
"""None multimodal_output when not finished returns None."""
transfer_manager = _make_transfer_manager()
request = _req("rid-none", finished=False)
payload = generator2tokenizer_async_chunk(
transfer_manager=transfer_manager,
pooling_output=None,
multimodal_output=None,
request=request,
)