[Refactor] Standardize data entry key names to {type}.{qualifier} format (#1829)

Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com>
Signed-off-by: dsinghvi <divyanshsinghvi@gmail.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Co-authored-by: Zhou Taichang <tzhouam@connect.ust.hk>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
authored and GitHub committed 2026-04-25 18:15:38 +08:00
1 parent 755a2d9d96
commit 2390350533
41 files changed
+1192 -569

No files matched your search

@@ -44,7 +44,7 @@ def test_forward_trims_context_on_exact_frame_boundaries():
out = model.forward(
input_ids=torch.arange(12, dtype=torch.long),
runtime_additional_information=[{"left_context_size": 2}],
runtime_additional_information=[{"meta": {"left_context_size": 2}}],
)
audio = out.multimodal_outputs["model_outputs"][0]
@@ -57,7 +57,7 @@ def test_forward_trims_trailing_padding_without_context():
out = model.forward(
input_ids=torch.arange(12, dtype=torch.long),
runtime_additional_information=[{"left_context_size": 0}],
runtime_additional_information=[{"meta": {"left_context_size": 0}}],
)
audio = out.multimodal_outputs["model_outputs"][0]
@@ -45,7 +45,7 @@ def test_dac_decoder_mixed_batch_empty_request_does_not_misalign_indices():
out = decoder.forward(
input_ids=torch.arange(20, dtype=torch.long),
runtime_additional_information=[{}, {"left_context_size": 1}],
runtime_additional_information=[{}, {"meta": {"left_context_size": 1}}],
)
audios = out.multimodal_outputs["model_outputs"]
@@ -58,8 +58,8 @@ def test_text2flow_supports_batched_source_outputs():
assert len(outputs) == 2
assert outputs[0]["prompt_token_ids"] == [1, 2, 3]
assert outputs[1]["prompt_token_ids"] == [4, 5]
assert outputs[0]["additional_information"]["prefix_ids"] == [10, 11]
assert outputs[1]["additional_information"]["prefix_ids"] == [20, 21]
assert outputs[0]["additional_information"]["ids"]["prompt"] == [10, 11]
assert outputs[1]["additional_information"]["ids"]["prompt"] == [20, 21]
def test_talker2code2wav_async_chunk_final_payload_uses_absolute_token_offset():
@@ -83,8 +83,8 @@ def test_talker2code2wav_async_chunk_final_payload_uses_absolute_token_offset():
)
assert payload is not None
assert payload["finished"].item() is True
assert payload["code_predictor_codes"] == [1, 2, 3]
assert payload["meta"]["finished"].item() is True
assert payload["codes"]["audio"] == [1, 2, 3]
assert payload["token_offset"] == 0
assert payload["left_context_size"] == 0
assert payload["req_id"] == ["rid-0"]
@@ -111,8 +111,8 @@ def test_talker2code2wav_async_chunk_emits_eof_when_finished_without_valid_codes
)
assert payload is not None
assert payload["code_predictor_codes"] == []
assert payload["finished"].item() is True
assert payload["codes"]["audio"] == []
assert payload["meta"]["finished"].item() is True
def test_talker2code2wav_async_chunk_does_not_reemit_without_new_tokens():
@@ -138,7 +138,7 @@ def test_talker2code2wav_async_chunk_does_not_reemit_without_new_tokens():
)
assert payload1 is not None
assert payload1["code_predictor_codes"] == [1, 2]
assert payload1["codes"]["audio"] == [1, 2]
assert payload1["token_offset"] == 0
assert payload2 is None
@@ -168,9 +168,9 @@ def test_talker2code2wav_async_chunk_waits_for_prelookahead_and_emits_cumulative
assert payload_pending is None
assert payload_ready is not None
assert payload_ready["code_predictor_codes"] == [1, 2, 3]
assert payload_ready["codes"]["audio"] == [1, 2, 3]
assert payload_ready["token_offset"] == 0
assert payload_ready["finished"].item() is False
assert payload_ready["meta"]["finished"].item() is False
def test_talker2code2wav_async_chunk_final_flush_uses_previous_token_offset():
@@ -197,12 +197,12 @@ def test_talker2code2wav_async_chunk_final_flush_uses_previous_token_offset():
)
assert payload_stream is not None
assert payload_stream["finished"].item() is False
assert payload_stream["code_predictor_codes"] == [3, 4, 5]
assert payload_stream["meta"]["finished"].item() is False
assert payload_stream["codes"]["audio"] == [3, 4, 5]
assert payload_stream["token_offset"] == 0
assert payload_final is not None
assert payload_final["finished"].item() is True
assert payload_final["code_predictor_codes"] == [3, 4, 5, 6]
assert payload_final["meta"]["finished"].item() is True
assert payload_final["codes"]["audio"] == [3, 4, 5, 6]
assert payload_final["token_offset"] == 2
@@ -233,7 +233,7 @@ def test_talker2code2wav_async_chunk_respects_prompt_token_pad_on_first_chunk():
assert payload_pending is None
assert payload_ready is not None
assert payload_ready["code_predictor_codes"] == [8, 9, 10, 11]
assert payload_ready["codes"]["audio"] == [8, 9, 10, 11]
assert payload_ready["token_offset"] == 0
@@ -260,8 +260,8 @@ def test_talker2code2wav_async_chunk_emits_terminal_eof_without_duplicate_audio(
)
assert payload_stream is not None
assert payload_stream["finished"].item() is False
assert payload_stream["code_predictor_codes"] == [3, 4]
assert payload_stream["meta"]["finished"].item() is False
assert payload_stream["codes"]["audio"] == [3, 4]
assert payload_final is not None
assert payload_final["finished"].item() is True
assert payload_final["code_predictor_codes"] == []
assert payload_final["meta"]["finished"].item() is True
assert payload_final["codes"]["audio"] == []
@@ -299,7 +299,7 @@ class TestAr2Diffusion:
def test_i2i_with_mm_output(self):
"""Test image-to-image with prior_token_image_ids from AR model."""
token_ids = list(range(1024)) + [16385]
mm_output = {"prior_token_image_ids": torch.tensor([1, 2, 3])}
mm_output = {"ids": {"prior_image": torch.tensor([1, 2, 3])}}
stage_list = [_stage_with_outputs([_source_output(token_ids, mm_output)])]
from PIL import Image
@@ -13,23 +13,23 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
def _sentinel():
return {"code_predictor_codes": [], "finished": torch.tensor(True, dtype=torch.bool)}
return {"codes": {"audio": []}, "meta": {"finished": torch.tensor(True, dtype=torch.bool)}}
def test_flush_remaining_codes_when_no_codes_accumulated_missing_request_id():
"""No entry for request_id: treat as empty, return finished sentinel with empty codes."""
tm = SimpleNamespace(code_prompt_token_ids={})
out = _flush_remaining_codes(tm, "missing", chunk_size=3, left_context_size=3)
assert out["code_predictor_codes"] == _sentinel()["code_predictor_codes"]
assert out["finished"].equal(_sentinel()["finished"])
assert out["codes"]["audio"] == _sentinel()["codes"]["audio"]
assert out["meta"]["finished"].item() is True
def test_flush_remaining_codes_when_no_codes_accumulated_empty_list():
"""Explicit empty accumulation list returns the same sentinel."""
tm = SimpleNamespace(code_prompt_token_ids={"r": []})
out = _flush_remaining_codes(tm, "r", chunk_size=3, left_context_size=3)
assert out["code_predictor_codes"] == []
assert out["finished"].item() is True
assert out["codes"]["audio"] == []
assert out["meta"]["finished"].item() is True
def test_flush_remaining_codes_partial_chunk_remaining():
@@ -41,8 +41,8 @@ def test_flush_remaining_codes_partial_chunk_remaining():
code_prompt_token_ids={"r": [[1], [2], [3], [4], [5], [6], [7]]},
)
out = _flush_remaining_codes(tm, "r", chunk_size=3, left_context_size=3)
assert out["finished"].item() is True
assert out["code_predictor_codes"] == [4, 5, 6, 7]
assert out["meta"]["finished"].item() is True
assert out["codes"]["audio"] == [4, 5, 6, 7]
def test_flush_remaining_codes_when_length_is_exact_multiple_of_chunk_size():
@@ -52,7 +52,7 @@ def test_flush_remaining_codes_when_length_is_exact_multiple_of_chunk_size():
)
out = _flush_remaining_codes(tm, "r", chunk_size=3, left_context_size=3)
# context_length = chunk_size = 3, end_index = min(6, 6) -> all 6
assert out["code_predictor_codes"] == [1, 2, 3, 4, 5, 6]
assert out["codes"]["audio"] == [1, 2, 3, 4, 5, 6]
@pytest.mark.parametrize(
@@ -74,5 +74,5 @@ def test_flush_remaining_codes_context_window_end_index(
tm = SimpleNamespace(code_prompt_token_ids={"r": accumulated})
out = _flush_remaining_codes(tm, "r", chunk_size=chunk_size, left_context_size=left_context)
expected_flat = list(range(length - expected_end_index, length))
assert out["code_predictor_codes"] == expected_flat
assert out["finished"].item() is True
assert out["codes"]["audio"] == expected_flat
assert out["meta"]["finished"].item() is True
@@ -19,7 +19,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
def _make_stage_list(codec_codes: torch.Tensor, request_id: str = "req-0"):
"""Build a minimal stage_list[0] with a single talker output carrying codec_codes."""
output = SimpleNamespace(multimodal_output={"code_predictor_codes": codec_codes})
output = SimpleNamespace(multimodal_output={"codes": {"audio": codec_codes}})
talker_output = SimpleNamespace(outputs=[output], request_id=request_id)
stage0 = SimpleNamespace(engine_outputs=[talker_output])
return [stage0]
@@ -55,7 +55,7 @@ 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={"audio_codes": torch.zeros((0,))},
pooling_output={"codes": {"audio": torch.zeros((0,))}},
request=_req(rid, finished=finished, initial_codec_chunk_frames=req_ic),
is_finished=finished,
)
@@ -65,7 +65,7 @@ def test_empty_returns_none():
tm = _tm()
p = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"audio_codes": torch.zeros((0,))},
pooling_output={"codes": {"audio": torch.zeros((0,))}},
request=_req("r", finished=False),
)
assert p is None
@@ -79,7 +79,8 @@ def test_eof_marker_when_finished_empty():
request=_req("r", finished=True),
is_finished=True,
)
assert p == {"code_predictor_codes": [], "finished": torch.tensor(True, dtype=torch.bool)}
assert p["codes"] == {"audio": []}
assert p["meta"]["finished"].item() is True
def test_flush_on_finish():
@@ -92,8 +93,8 @@ def test_flush_on_finish():
is_finished=True,
)
assert p is not None
assert p["finished"] is True
assert len(p["code_predictor_codes"]) == _Q * 24
assert p["meta"]["finished"].item() is True
assert len(p["codes"]["audio"]) == _Q * 24
_CASES = [
@@ -158,8 +159,8 @@ def test_streaming_phases(config, n_frames, finished, expected):
else:
exp_ctx, exp_window = expected
assert payload is not None
assert payload["left_context_size"] == exp_ctx
assert len(payload["code_predictor_codes"]) == _Q * exp_window
assert payload["meta"]["left_context_size"] == exp_ctx
assert len(payload["codes"]["audio"]) == _Q * exp_window
def test_dynamic_ic_adapts_to_load():
@@ -169,14 +170,14 @@ def test_dynamic_ic_adapts_to_load():
# Low load (1/8) -> IC=2 -> emit at 2
p1 = _call(tm, "r", n_frames=2)
assert p1 is not None
assert len(p1["code_predictor_codes"]) == _Q * 2
assert len(p1["codes"]["audio"]) == _Q * 2
# High load: add 4 others -> active=5/8 -> IC=8 -> emit at 8
for i in range(4):
tm.code_prompt_token_ids[f"other-{i}"] = [[0]]
p2 = _call(tm, "r", n_frames=8)
assert p2 is not None
assert len(p2["code_predictor_codes"]) == _Q * 8
assert len(p2["codes"]["audio"]) == _Q * 8
# Requests past initial phase still count in load factor
tm2 = _tm(max_num_seqs=4)
@@ -185,7 +186,7 @@ def test_dynamic_ic_adapts_to_load():
# active=4/4=1.0 -> IC=16
p3 = _call(tm2, "new", n_frames=16)
assert p3 is not None
assert len(p3["code_predictor_codes"]) == _Q * 16
assert len(p3["codes"]["audio"]) == _Q * 16
def test_ic_load_change_mid_request():
@@ -206,6 +207,7 @@ def test_ic_load_change_mid_request():
assert _call(tm, "r", n_frames=27) is None
p3 = _call(tm, "r", n_frames=49)
assert p3 is not None
assert p3["meta"]["left_context_size"] == 24
# A *new* request under high load gets IC=16 (not IC=2).
# Frame 2 would emit under IC=2 but must hold under IC=16.
@@ -253,14 +255,14 @@ def test_first_streaming_chunk_prepends_ref_code_context():
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"audio_codes": torch.zeros((0,)), "ref_code": ref_code},
pooling_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
assert payload is not None
assert payload["left_context_size"] == 2
assert len(payload["code_predictor_codes"]) == _Q * 12
assert payload["meta"]["left_context_size"] == 2
assert len(payload["codes"]["audio"]) == _Q * 12
def test_ref_code_context_applies_to_all_streaming_chunks():
@@ -274,15 +276,15 @@ def test_ref_code_context_applies_to_all_streaming_chunks():
payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"audio_codes": torch.zeros((0,)), "ref_code": ref_code},
pooling_output={"codes": {"audio": torch.zeros((0,)), "ref": ref_code}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
assert payload is not None
# ref_code (2 frames) prepended as left context on second chunk too
assert payload["left_context_size"] == 10 + 2
assert len(payload["code_predictor_codes"]) == _Q * (20 + 2)
assert payload["meta"]["left_context_size"] == 10 + 2
assert len(payload["codes"]["audio"]) == _Q * (20 + 2)
def test_ref_code_context_can_be_buffered_before_first_emit():
@@ -292,7 +294,7 @@ def test_ref_code_context_can_be_buffered_before_first_emit():
first_payload = talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"audio_codes": torch.tensor([[1, 2, 3, 4]]), "ref_code": ref_code},
pooling_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,
)
@@ -302,22 +304,22 @@ def test_ref_code_context_can_be_buffered_before_first_emit():
for _ in range(8):
talker2code2wav_async_chunk(
transfer_manager=tm,
pooling_output={"audio_codes": torch.tensor([[1, 2, 3, 4]])},
pooling_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={"audio_codes": torch.tensor([[1, 2, 3, 4]])},
pooling_output={"codes": {"audio": torch.tensor([[1, 2, 3, 4]])}},
request=_req(rid, finished=False, initial_codec_chunk_frames=10),
is_finished=False,
)
assert payload is not None
# ref_code (2 frames) is kept (not popped) for subsequent chunks
assert payload["left_context_size"] == 2
assert len(payload["code_predictor_codes"]) == _Q * 12
assert payload["meta"]["left_context_size"] == 2
assert len(payload["codes"]["audio"]) == _Q * 12
assert rid in tm.request_payload
@@ -332,7 +334,7 @@ def test_non_async_processor_prepends_ref_code_and_sets_trim_context():
dtype=torch.long,
)
output = SimpleNamespace(
multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code},
multimodal_output={"codes": {"audio": audio_codes, "ref": ref_code}},
token_ids=list(range(3)),
)
stage = SimpleNamespace(
@@ -343,7 +345,7 @@ def test_non_async_processor_prepends_ref_code_and_sets_trim_context():
assert len(prompts) == 1
prompt = prompts[0]
assert prompt["additional_information"] == {"left_context_size": [2]}
assert prompt["additional_information"] == {"meta": {"left_context_size": [2]}}
assert prompt["prompt_token_ids"] == [
9,
8,
@@ -377,7 +379,7 @@ def test_non_async_processor_filters_out_of_range_codec_values():
dtype=torch.long,
)
output = SimpleNamespace(
multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code},
multimodal_output={"codes": {"audio": audio_codes, "ref": ref_code}},
token_ids=list(range(4)),
)
stage = SimpleNamespace(
@@ -390,4 +392,4 @@ def test_non_async_processor_filters_out_of_range_codec_values():
prompt = prompts[0]
# Only ref_code (1 frame) + 2 valid frames = 3 frames * 4 quantizers = 12 codes
assert len(prompt["prompt_token_ids"]) == 4 * 3
assert prompt["additional_information"] == {"left_context_size": [1]}
assert prompt["additional_information"] == {"meta": {"left_context_size": [1]}}
@@ -126,8 +126,8 @@ def test_flush_tail_when_finished():
)
assert payload is not None
assert payload["finished"].item() is True
codes = payload["code_predictor_codes"]
assert payload["meta"]["finished"].item() is True
codes = payload["codes"]["audio"]
# Format: [ctx_frames, context_length, ...flat_codes]
assert len(codes) >= 2 # At least ctx_frames + context_length header
ctx_frames = codes[0]
@@ -149,10 +149,8 @@ def test_eof_marker_when_finished_with_no_frames():
request=request,
)
assert payload == {
"code_predictor_codes": [],
"finished": torch.tensor(True, dtype=torch.bool),
}
assert payload["codes"] == {"audio": []}
assert payload["meta"]["finished"].item() is True
def test_normal_chunk_emission():
@@ -176,7 +174,7 @@ def test_normal_chunk_emission():
# A chunk should be emitted
assert payload is not None
codes = payload["code_predictor_codes"]
codes = payload["codes"]["audio"]
ctx_frames = codes[0]
context_length = codes[1]
assert ctx_frames == 20 # 25 - 5(chunk_size_at_begin)
@@ -203,7 +201,7 @@ def test_small_initial_chunks():
)
assert payload is not None
codes = payload["code_predictor_codes"]
codes = payload["codes"]["audio"]
ctx_frames = codes[0]
context_length = codes[1]
assert ctx_frames == 0
@@ -252,7 +250,7 @@ def test_context_handling_format():
)
assert payload is not None
codes = payload["code_predictor_codes"]
codes = payload["codes"]["audio"]
# First two elements are ctx_frames and context_length
ctx_frames = codes[0]
context_length = codes[1]