tests

tests/cossim_common.py

@@ -58,6 +58,13 @@ from qwen_tts.core.models.modeling_qwen3_tts      import Qwen3TTSForConditionalG
 from qwen_tts.core.models.configuration_qwen3_tts import Qwen3TTSConfig
 from qwen_tts.core.models.processing_qwen3_tts    import Qwen3TTSProcessor
 from transformers import AutoConfig, AutoModel, AutoProcessor
+from transformers.utils import logging as hf_logging
+
+# Silence the GenerationConfig validator that warns "flags are not valid
+# and may be ignored" for temperature / top_k / top_p when do_sample=False.
+# Those flags ride along inside the checkpoint generation_config and the
+# greedy path drops them on purpose, the warning is just noise here.
+hf_logging.set_verbosity_error()
 
 # Register the Qwen3-TTS classes once per process. Calling twice raises a
 # ValueError inside transformers, hence the guard.
@@ -378,3 +385,14 @@ def compare_exact_i32(name, dump_cpp, dump_pt, label):
     pct  = 100.0 * float(np.mean(ai[:n] == bi[:n]))
     print(f"[Cossim] {label} exact: {pct:.2f}% ({n} values)")
     return pct
+
+# Greedy generation kwargs shared by every cossim script. do_sample=False
+# alone selects argmax, top_k / top_p / temperature are intentionally
+# omitted because GenerationConfig flags them as "not valid" warnings when
+# do_sample=False. The subtalker_* keys are custom kwargs forwarded to the
+# talker forward, the talker validator is bypassed by the script main.
+GEN_KWARGS_GREEDY = dict(
+    do_sample             = False,
+    subtalker_dosample    = False,
+    repetition_penalty    = 1.0,
+)

tests/debug-base-cossim.py

@@ -87,31 +87,12 @@ def main():
     # the strict validator. Disable it on the talker only.
     model.talker._validate_model_kwargs = lambda *a, **k: None
 
-    # Greedy hardcoded : argmax on both talker c0 and code predictor sub
-    # codes. Stochastic mode is not exercised here because the F32 drift
-    # between torch CUDA cuBLAS and ggml CUDA matmul on Qwen3 norm_w
-    # inflated activations propagates through the FFN and flips multinomial
-    # picks in flat distributions, breaking bit exactness. Argmax is robust
-    # to that drift, so greedy gives 100 percent CodesFull match and
-    # validates the full forward + sampling chain.
-    gen_kwargs = dict(
-        do_sample             = False,
-        top_k                 = 1,
-        top_p                 = 1.0,
-        temperature           = 1.0,
-        subtalker_dosample    = False,
-        subtalker_top_k       = 1,
-        subtalker_top_p       = 1.0,
-        subtalker_temperature = 1.0,
-        repetition_penalty    = 1.0,
-    )
-
     talker_codes_list, _ = model.generate(
         input_ids=[input_ids],
         languages=[args.lang],
         non_streaming_mode=True,
         max_new_tokens=args.max_new_tokens,
-        **gen_kwargs,
+        **cc.GEN_KWARGS_GREEDY,
     )
     codes = talker_codes_list[0]
     print(f"[Python] Codes shape: {tuple(codes.shape)} (T_frames, num_code_groups)")

tests/debug-clone-cossim.py

@@ -147,18 +147,6 @@ def main():
     # the strict validator. Disable it on the talker only.
     model.talker._validate_model_kwargs = lambda *a, **k: None
 
-    gen_kwargs = dict(
-        do_sample             = False,
-        top_k                 = 1,
-        top_p                 = 1.0,
-        temperature           = 1.0,
-        subtalker_dosample    = False,
-        subtalker_top_k       = 1,
-        subtalker_top_p       = 1.0,
-        subtalker_temperature = 1.0,
-        repetition_penalty    = 1.0,
-    )
-
     # voice_clone_prompt dict mirrors what _prompt_items_to_voice_clone_prompt
     # builds for a single ICL prompt item : ref_code is the [T_codec, K]
     # tensor, ref_spk_embedding is the [hidden] tensor, x_vector_only=False
@@ -177,7 +165,7 @@ def main():
         languages=[args.lang],
         non_streaming_mode=False,
         max_new_tokens=args.max_new_tokens,
-        **gen_kwargs,
+        **cc.GEN_KWARGS_GREEDY,
     )
     codes = talker_codes_list[0]
     print(f"[Python] Codes shape: {tuple(codes.shape)} (T_frames, num_code_groups)")

tests/debug-customvoice-cossim.py

@@ -113,18 +113,6 @@ def main():
     # the strict validator. Disable it on the talker only.
     model.talker._validate_model_kwargs = lambda *a, **k: None
 
-    gen_kwargs = dict(
-        do_sample             = False,
-        top_k                 = 1,
-        top_p                 = 1.0,
-        temperature           = 1.0,
-        subtalker_dosample    = False,
-        subtalker_top_k       = 1,
-        subtalker_top_p       = 1.0,
-        subtalker_temperature = 1.0,
-        repetition_penalty    = 1.0,
-    )
-
     talker_codes_list, _ = model.generate(
         input_ids=[input_ids],
         instruct_ids=instruct_ids_arg,
@@ -132,7 +120,7 @@ def main():
         speakers=[args.speaker],
         non_streaming_mode=True,
         max_new_tokens=args.max_new_tokens,
-        **gen_kwargs,
+        **cc.GEN_KWARGS_GREEDY,
     )
     codes = talker_codes_list[0]
     print(f"[Python] Codes shape: {tuple(codes.shape)} (T_frames, num_code_groups)")

tests/debug-tts-cossim.py

@@ -103,25 +103,13 @@ def main():
     # the strict validator. Disable it on the talker only.
     model.talker._validate_model_kwargs = lambda *a, **k: None
 
-    gen_kwargs = dict(
-        do_sample             = False,
-        top_k                 = 1,
-        top_p                 = 1.0,
-        temperature           = 1.0,
-        subtalker_dosample    = False,
-        subtalker_top_k       = 1,
-        subtalker_top_p       = 1.0,
-        subtalker_temperature = 1.0,
-        repetition_penalty    = 1.0,
-    )
-
     talker_codes_list, _ = model.generate(
         input_ids=[input_ids],
         instruct_ids=[instruct_ids],
         languages=[args.lang],
         non_streaming_mode=True,
         max_new_tokens=args.max_new_tokens,
-        **gen_kwargs,
+        **cc.GEN_KWARGS_GREEDY,
     )
     codes = talker_codes_list[0]
     print(f"[Python] Codes shape: {tuple(codes.shape)} (T_frames, num_code_groups)")