Upload README.md with huggingface_hub

README.md

@@ -24,7 +24,7 @@ token_embeddings     input_ids      (1, L)         int64 ->  rows      (1, L, 10
 text_model           rows, positions, slots               ->  logits   (1, L, 151936)
 ```
 
-- **File**: `qwen3_asr_0_6b_xnnpack_8da4w.pte` — **697.5 MB**, fifteen methods
+- **File**: `qwen3_asr_0_6b_xnnpack_8da4w.pte` — **697.0 MB**, fifteen methods
 - **Source**: [Qwen/Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B) — 782.4M parameters
 - **License**: apache-2.0
 - **Languages**: 30, auto-detected — the model emits `language <Name>` before the text
@@ -88,7 +88,9 @@ tipping rather than silence gradually swamping the signal. The errors look like
 
 **Hence the ladder.** Thirteen encoder methods cost **3.8 MB** in total, because
 methods in one `.pte` share their constants: one window is 3753.4 MB in fp32 and
-thirteen are 3757.2 MB. There is no reason to make a caller over-pad.
+thirteen are 3757.2 MB. There is no reason to make a caller over-pad. (Those two fp32
+figures are from the pre-rewrite graph; the rewrite moved the published 8da4w file by
+0.5 MB, and the +3.8 MB it is making a point about is unaffected.)
 
 ## Verification
 
@@ -99,7 +101,7 @@ and window. Nine clips, Japanese / Portuguese / Russian, windows 2–4 s:
 | bundle | worst CER | ignoring case and terminal punctuation |
 |---|---|---|
 | fp32 (3757.2 MB, not published) | **0.0000** | 0.0000 |
-| 8da4w (697.5 MB, this file) | 0.1333 | **0.0667** |
+| 8da4w (697.0 MB, this file) | 0.1333 | **0.0667** |
 
 Both arms run at the same window on purpose, so this measures the conversion and not
 the window choice, which is measured separately above.
@@ -140,16 +142,28 @@ reproduces the fp32 model, and nothing about word error rate on real speakers.
 Measured standalone at the 30-second window, swapping only the audio half and reading
 the transcript:
 
-| build | size (MB) | Mac ms* | corr vs eager | worst CER |
-|---|---|---|---|---|
-| XNNPACK fp32 | 746.1 | 170.0 | 0.999998 | 0.0000 |
-| XNNPACK fp16 | 381.9 | 320.5 | 0.999991 | 0.0000 |
-| XNNPACK int8 | 200.7 | 149.4 | 0.998460 | **0.0000** |
-| Core ML | 374.1 | **40.9** | 0.973094 | **0.0690** |
-
-\*Mac arm64, median of 5 — a reference point for relative cost, not a device number.
-Torch eager fp32 on the same machine is 558.1 ms, so the XNNPACK build is already
-3.3x faster than eager.
+| build | size (MB) | Mac ms* | XNNPACK takes | corr vs eager | worst CER† |
+|---|---|---|---|---|---|
+| XNNPACK fp32 | 746.0 | 143.6 | 71.9% | 0.999999 | 0.0000 |
+| XNNPACK fp16 | 381.7 | 237.6 | 64.5% | 0.999995 | 0.0000 |
+| XNNPACK int8 | 200.6 | 123.5 | 78.0% | 0.998266 | **0.0000** |
+| Core ML | 374.1 | **40.6** | 100% | 0.972872 | **0.0690** |
+
+\*Mac arm64, median of 10 — a reference point for relative cost, not a device number.
+Torch eager fp32 on the same machine is 600.6 ms, so the XNNPACK build is 4.2x faster
+than eager.
+
+†The CER column was measured on the pre-rewrite graph (below). The rewrite agrees with
+it to 1.3e-06 at the tensor level and leaves every transcript in this card unchanged,
+but the transcripts behind that column were not re-read.
+
+**Those speeds moved because the windowed attention stopped calling
+`F.scaled_dot_product_attention`.** It decomposes to `_safe_softmax`, whose guard for
+fully-masked rows is six operations XNNPACK cannot take — and the call has no mask and
+`is_causal=False`, so the guard can never fire. There are 18 layers x 4 windows of it at
+30 seconds. Writing the softmax out took the fp32 tower from **55.8% delegated at
+170.0 ms to 71.9% at 143.6 ms**, and int8 from 149.4 ms to 123.5 ms, agreeing with the
+old graph to 7.1e-07.
 
 Two things worth taking from that table. **int8 on the audio tower is free** in the
 only unit that matters, which is why the shipped bundle uses it there. And **the Core