tts4_int4: set measured steps/cfg

tts4_int4/manifest.json

@@ -1,45 +1,46 @@
 {
-  "graphs": {
-    "dit_step": "dit_step.onnx",
-    "vocoder": "vocoder.onnx"
-  },
-  "vocoder_tail": {
-    "why": "this vocos variant ends in an ISTFTHead (complex spectrogram); ONNX has no complex tensors, so the graph stops at the linear head",
-    "planes": "[B, 1026, T] = concat(mag_pre, phase)",
-    "browser_tail": [
-      "mag = min(exp(planes[:513]), 1e2)",
-      "phase = planes[513:]",
-      "spec = mag * (cos(phase) + i * sin(phase))   # [513, T]",
-      "wav = istft(spec, n_fft=1024, hop=256, win=1024, hann, center=True)"
-    ],
-    "reference": "web/src/audio/istft.js (pinned by tests/test_istft.py)",
-    "samples_per_frame": 256
-  },
-  "dtype": "float16",
-  "opset": 20,
-  "mel": {
-    "target_sample_rate": 24000,
-    "n_mel_channels": 100,
-    "hop_length": 256,
-    "n_fft": 1024,
-    "win_length": 1024,
-    "mel_spec_type": "vocos"
-  },
-  "vocab": {
-    "file": "vocab.txt",
-    "size": 2545,
-    "note": "ASCII + numbered pinyin (ai3, ni3); NO CJK glyphs, so Chinese text must be converted to pinyin with tones before lookup"
-  },
-  "sampler": {
-    "solver": "euler",
-    "steps": 32,
-    "sway_sampling_coef": -1.0,
-    "schedule": "t = linspace(0, 1, steps + 1); t = t + sway * (cos(pi/2 * t) - 1 + t); step y += dt * v",
-    "cfg_strength": 2.0,
-    "cfg_note": "run the step graph twice (cond and text, then cond=0 and text=0) and combine pred + (pred - null) * cfg_strength; the uncond call differs slightly from the repo's internal drop_text (filler-token embedding vs learned zeros), so set cfg_strength=0 for exact parity"
-  },
-  "vocoder_chunk_frames": 375,
-  "vocoder_chunk_note": "chunk long mel (375 frames ~ 4 s at hop 256) rather than one huge pass: the backbone is convolutional, so chunking changes the boundary only",
-  "sample_rate": 24000,
-  "source_ckpt": "/root/WebASMR/models/f5tts_4step.pt"
+ "graphs": {
+  "dit_step": "dit_step.onnx",
+  "vocoder": "vocoder.onnx"
+ },
+ "vocoder_tail": {
+  "why": "this vocos variant ends in an ISTFTHead (complex spectrogram); ONNX has no complex tensors, so the graph stops at the linear head",
+  "planes": "[B, 1026, T] = concat(mag_pre, phase)",
+  "browser_tail": [
+   "mag = min(exp(planes[:513]), 1e2)",
+   "phase = planes[513:]",
+   "spec = mag * (cos(phase) + i * sin(phase))   # [513, T]",
+   "wav = istft(spec, n_fft=1024, hop=256, win=1024, hann, center=True)"
+  ],
+  "reference": "web/src/audio/istft.js (pinned by tests/test_istft.py)",
+  "samples_per_frame": 256
+ },
+ "dtype": "float16",
+ "opset": 20,
+ "mel": {
+  "target_sample_rate": 24000,
+  "n_mel_channels": 100,
+  "hop_length": 256,
+  "n_fft": 1024,
+  "win_length": 1024,
+  "mel_spec_type": "vocos"
+ },
+ "vocab": {
+  "file": "vocab.txt",
+  "size": 2545,
+  "note": "ASCII + numbered pinyin (ai3, ni3); NO CJK glyphs, so Chinese text must be converted to pinyin with tones before lookup"
+ },
+ "sampler": {
+  "solver": "euler",
+  "steps": 4,
+  "sway_sampling_coef": -1.0,
+  "schedule": "t = linspace(0, 1, steps + 1); t = t + sway * (cos(pi/2 * t) - 1 + t); step y += dt * v",
+  "cfg_strength": 0.0,
+  "cfg_note": "run the step graph twice (cond and text, then cond=0 and text=0) and combine pred + (pred - null) * cfg_strength; the uncond call differs slightly from the repo's internal drop_text (filler-token embedding vs learned zeros), so set cfg_strength=0 for exact parity",
+  "measured_note": "4 steps, CFG 0: reflow-distilled student. CER 0.597 vs 0.775 for the teacher at 4:0, still short of the 16-step 0.097. Experimental."
+ },
+ "vocoder_chunk_frames": 375,
+ "vocoder_chunk_note": "chunk long mel (375 frames ~ 4 s at hop 256) rather than one huge pass: the backbone is convolutional, so chunking changes the boundary only",
+ "sample_rate": 24000,
+ "source_ckpt": "/root/WebASMR/models/f5tts_4step.pt"
 }