[mudler-agent]

What

Fix a numerical mismatch with NeMo for the Nemotron unified ASR checkpoint on clips whose trimmed mel has 8k + 1 frames (for example 625 frames, a clip of 100000 to 100159 samples).

Cause

The unified frontend trims the mel to NeMo's valid length, so every mel frame is valid. The offline subsampling still assumed the untrimmed convention (valid length = T - 1, one trailing pad frame from center padding). The three stride-2 convolutions map a length n to ceil(n / 2), so T and T - 1 give different valid output lengths when T = 8k + 1. There the last output frame was masked to zero (78 valid frames instead of 79), which changed the encoder output and dropped the last tokens.

Change

Subsampling::offline_entry_valid(T) returns T when the checkpoint trims the mel (parakeet.unified.valid_length) and T - 1 otherwise. The single, batched and tiled offline paths and valid_out_len use it. Other models keep T - 1 and are unchanged. Streaming paths pass explicit lengths and are not affected.

Verification

Unified f32 GGUF, English, against NeMo (encoder max abs difference, tokens):

Samples	Mel frames	Before	After
100000	625	0.115, 42 vs 45 tokens	1.8e-7, 45/45
101280	633	not run	4.8e-7, 43/43
98720	617	not run	4.8e-7, 45/45
100160	626 (control)	not run	4.8e-7, 45/45

• New checkpoint-free test of the valid output length with the gate on and off (test_unified_synthetic).
• ctest -LE model: all pass on a CPU build.
• Not run: German, f16, GPU.

This also explains the two failing runs of test_unified_stream seen with a window of exactly 625 frames.

🤖 Generated with Claude Code