Profile pinned S25 and X Elite vocoders on Qualcomm hardware

benchmarks/edge_harness/experiments/audit_qwen_tts_real_code_device.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+"""Bind a generated-code S25 waveform to its pinned TFLite target and CPU references."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+from pathlib import Path
+
+import numpy as np
+
+
+def sha256(path: Path) -> str:
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def waveform(path: Path, key: str) -> np.ndarray:
+    with np.load(path, allow_pickle=False) as data:
+        if data.files != [key]:
+            raise ValueError(f"{path}: expected only {key}")
+        value = np.asarray(data[key])
+    if value.shape != (1, 48000) or value.dtype != np.float32 or not np.isfinite(value).all():
+        raise ValueError(f"{path}: invalid 25-frame waveform {value.shape}/{value.dtype}")
+    return value
+
+
+def compare(reference: np.ndarray, observed: np.ndarray) -> dict:
+    a = reference.astype(np.float64).ravel()
+    b = observed.astype(np.float64).ravel()
+    error = float(np.linalg.norm(a - b))
+    norm = float(np.linalg.norm(a))
+    return {
+        "relative_l2": error / norm,
+        "snr_db": 20 * math.log10(norm / error) if error else None,
+        "max_abs": float(np.max(np.abs(a - b))),
+        "reference_rms": float(np.sqrt(np.mean(a * a))),
+        "device_rms": float(np.sqrt(np.mean(b * b))),
+        "device_saturated_fraction_abs_ge_0_999": float(np.mean(np.abs(b) >= 0.999)),
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--evidence-dir", type=Path, required=True)
+    parser.add_argument("--pinned-compile", type=Path, required=True)
+    parser.add_argument("--pinned-export", type=Path, required=True)
+    args = parser.parse_args()
+    p = args.evidence_dir
+    local = json.loads((p / "report.json").read_text())
+    dataset = json.loads((p / "dataset_submission.json").read_text())
+    submitted = json.loads((p / "inference_submission.json").read_text())
+    inference = json.loads((p / "inference_report.json").read_text())
+    compiled = json.loads(args.pinned_compile.read_text())
+    exported = json.loads(args.pinned_export.read_text())
+    if (local["model_revision"] != exported["model_snapshot_revision"]
+            or local["decoder_weight_sha256"] != exported["model_weight_sha256"]
+            or local["onnx_sha256"] != exported["onnx_sha256"]):
+        raise ValueError("generated codes and compiled source are not from the pinned export")
+    if (local["files"]["fixture"]["sha256"] != dataset["fixture_sha256"]
+            or dataset["input_shape"] != [1, 512, 97]
+            or sha256(p / "fixture.npz") != dataset["fixture_sha256"]):
+        raise ValueError("device input differs from real generated-code fixture")
+    if (submitted["job_id"] != inference["job_id"]
+            or inference["status"] != "SUCCESS"
+            or inference["model_id"] != compiled["target_model_id"]
+            or inference["input_dataset_id"] != dataset["dataset_id"]
+            or inference["options"] != "--compute_unit gpu"
+            or inference["device"] != compiled["device"]
+            or inference["device"]["name"] != "Samsung Galaxy S25"):
+        raise ValueError("device result does not bind to pinned S25 GPU target")
+    for name in ("eager", "ort", "full_reference"):
+        entry = local["files"][name]
+        if sha256(p / entry["filename"]) != entry["sha256"]:
+            raise ValueError(f"local {name} reference changed")
+    cpu = waveform(p / "ort_output.npz", "wav")
+    eager = waveform(p / "eager_output.npz", "wav")
+    full = waveform(p / "full_reference_segment.npz", "wav")
+    hosted = waveform(p / "device_output.npz", "output_0__0")
+    comparison = compare(cpu, hosted)
+    gross_failure = (comparison["relative_l2"] >= 1.0
+                     or comparison["device_saturated_fraction_abs_ge_0_999"] >= 0.1)
+    report = {
+        "scope": "one generated-code 25-frame vocoder window on hosted S25; no device-local talker or complete stream",
+        "status": "component_gross_numeric_failure" if gross_failure else "component_numeric_measured_no_speech_quality_gate",
+        "device": inference["device"],
+        "source_model_id": compiled["source_model_id"],
+        "target_model_id": compiled["target_model_id"],
+        "input_dataset_id": dataset["dataset_id"],
+        "jobs": {"compile": compiled["job_id"], "inference": inference["job_id"]},
+        "files": {name: {"sha256": sha256(p / filename), "bytes": (p / filename).stat().st_size}
+                  for name, filename in {"fixture": "fixture.npz", "device": "device_output.npz",
+                                         "ort": "ort_output.npz", "eager": "eager_output.npz",
+                                         "full": "full_reference_segment.npz"}.items()},
+        "ort_cpu_vs_device": comparison,
+        "eager_cpu_vs_device": compare(eager, hosted),
+        "full_fp32_decoder_segment_vs_device": compare(full, hosted),
+        "gross_numeric_failure_screen": {
+            "failed": gross_failure,
+            "criteria": "relative L2 >= 1.0 or >=10% saturated samples; not a speech-quality tolerance",
+        },
+        "limits": [
+            "One generated-code window is not a speech listening or intelligibility test.",
+            "Talker, predictor, co-resident handoff, state, cancellation and memory admission did not run on S25.",
+            "The separate placement profile used Workbench-generated same-shape input, not this fixture.",
+        ],
+    }
+    (p / "audit_report.json").write_text(json.dumps(report, indent=2) + "\n")
+    print(json.dumps({"status": report["status"], "ort_cpu_vs_device": comparison}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/edge_harness/experiments/probe_qwen_tts_real_code_window.py

@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+"""Capture one generated CustomVoice code stream and audit a 25-frame export.
+
+Set PYTHONPATH to include the pinned qwen_tts source checkout when invoking.
+The hosted device is not used here; this checks a real-code local window and
+its truncation against the full FP32 decoder before a device quality attempt.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import inspect
+import json
+import math
+import os
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+import torch
+from qwen_tts import Qwen3TTSModel
+from qwen_tts.core.tokenizer_12hz.configuration_qwen3_tts_tokenizer_v2 import (
+    Qwen3TTSTokenizerV2Config,
+)
+from qwen_tts.core.tokenizer_12hz.modeling_qwen3_tts_tokenizer_v2 import (
+    Qwen3TTSTokenizerV2Decoder,
+)
+from safetensors.torch import load_file
+
+
+def load_decoder(model: Path) -> Qwen3TTSTokenizerV2Decoder:
+    config = Qwen3TTSTokenizerV2Config.from_pretrained(str(model), subfolder="speech_tokenizer")
+    decoder = Qwen3TTSTokenizerV2Decoder._from_config(config.decoder_config)
+    state = load_file(str(model / "speech_tokenizer" / "model.safetensors"))
+    decoder_state = {key[len("decoder."):]: value for key, value in state.items()
+                     if key.startswith("decoder.")}
+    missing, unexpected = decoder.load_state_dict(decoder_state, strict=False)
+    if missing or unexpected:
+        raise RuntimeError(f"decoder state mismatch: {len(missing)} missing, {len(unexpected)} unexpected")
+    decoder.eval().float()
+    if hasattr(decoder, "precompute_snake_caches"):
+        decoder.precompute_snake_caches()
+    return decoder
+
+
+def decode_quantized(decoder: Qwen3TTSTokenizerV2Decoder, quantized: torch.Tensor) -> torch.Tensor:
+    hidden = decoder.pre_conv(quantized).transpose(1, 2)
+    hidden = decoder.pre_transformer(inputs_embeds=hidden).last_hidden_state
+    hidden = hidden.permute(0, 2, 1)
+    for blocks in decoder.upsample:
+        for block in blocks:
+            hidden = block(hidden)
+    for block in decoder.decoder:
+        hidden = block(hidden)
+    return hidden.clamp(-1, 1).reshape(1, -1)
+
+
+def sha256(path: Path) -> str:
+    with path.open("rb") as stream:
+        return hashlib.file_digest(stream, "sha256").hexdigest()
+
+
+def compare(reference: np.ndarray, observed: np.ndarray) -> dict:
+    a = reference.astype(np.float64).ravel()
+    b = observed.astype(np.float64).ravel()
+    error = float(np.linalg.norm(a - b))
+    norm = float(np.linalg.norm(a))
+    return {
+        "relative_l2": error / norm,
+        "snr_db": 20 * math.log10(norm / error) if error else None,
+        "max_abs": float(np.max(np.abs(a - b))),
+        "reference_rms": float(np.sqrt(np.mean(a * a))),
+        "observed_rms": float(np.sqrt(np.mean(b * b))),
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model", required=True, type=Path)
+    parser.add_argument("--onnx", required=True, type=Path)
+    parser.add_argument("--output-dir", required=True, type=Path)
+    parser.add_argument("--text", default=(
+        "The small red square is centered on the white card. I can see its "
+        "straight edges clearly, and the surrounding background is bright "
+        "and uncluttered."
+    ))
+    parser.add_argument("--max-new-tokens", type=int, default=144)
+    parser.add_argument("--threads", type=int, default=8)
+    args = parser.parse_args()
+    if args.max_new_tokens < 97 or args.threads <= 0:
+        parser.error("at least 97 max new tokens and positive threads required")
+    os.environ["HF_HUB_OFFLINE"] = "1"
+    os.environ["TRANSFORMERS_OFFLINE"] = "1"
+    torch.set_num_threads(args.threads)
+
+    wrapper = Qwen3TTSModel.from_pretrained(
+        str(args.model.resolve()), device_map="cpu", dtype=torch.bfloat16,
+        attn_implementation="sdpa", local_files_only=True,
+    )
+    captured: list[torch.Tensor] = []
+    original_decode = wrapper.model.speech_tokenizer.decode
+
+    def capture(items: list[dict], *extra: object, **kwargs: object):
+        captured.extend(item["audio_codes"].detach().cpu().clone() for item in items)
+        return original_decode(items, *extra, **kwargs)
+
+    wrapper.model.speech_tokenizer.decode = capture
+    torch.manual_seed(42)
+    wavs, sample_rate = wrapper.generate_custom_voice(
+        text=args.text, language="English", speaker="Ryan",
+        max_new_tokens=args.max_new_tokens, do_sample=True,
+        non_streaming_mode=True,
+    )
+    if len(captured) != 1 or len(wavs) != 1 or sample_rate != 24000:
+        raise RuntimeError("expected one generated code stream and 24 kHz waveform")
+    codes = captured[0]
+    if codes.ndim != 2 or codes.shape[0] < 97 or codes.shape[1] != 16:
+        raise RuntimeError(f"expected >=97 frames of 16-code speech, got {tuple(codes.shape)}")
+    generated = np.asarray(wavs[0], dtype=np.float32).ravel()
+    if len(generated) != codes.shape[0] * 1920 or not np.isfinite(generated).all():
+        raise RuntimeError("complete request has invalid audio length or samples")
+
+    del wrapper
+    decoder = load_decoder(args.model)
+    window_codes = codes[:97].T.unsqueeze(0).contiguous().long()
+    full_codes = codes.T.unsqueeze(0).contiguous().long()
+    with torch.inference_mode():
+        quantized = decoder.quantizer.decode(window_codes).detach().float().numpy()
+        eager = decode_quantized(decoder, torch.from_numpy(quantized))[:, -25 * 1920:].detach().numpy()
+        full = decoder(full_codes).reshape(1, -1).detach().numpy()
+    if quantized.shape != (1, 512, 97) or eager.shape != (1, 48000):
+        raise RuntimeError("real-code window violates fixed export shape")
+    full_segment = full[:, 72 * 1920 : 97 * 1920]
+    if full_segment.shape != eager.shape:
+        raise RuntimeError("full decoder output does not align with window")
+    options = ort.SessionOptions()
+    options.intra_op_num_threads = args.threads
+    session = ort.InferenceSession(str(args.onnx), sess_options=options,
+                                   providers=["CPUExecutionProvider"])
+    observed = np.asarray(session.run(["wav"], {"quantized": quantized})[0])
+    if observed.shape != eager.shape or not np.isfinite(observed).all():
+        raise RuntimeError("ONNX output violates waveform contract")
+
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    files = {
+        "codes": ("codes.npz", {"codes": codes.numpy()}),
+        "fixture": ("fixture.npz", {"quantized": quantized}),
+        "eager": ("eager_output.npz", {"wav": eager}),
+        "ort": ("ort_output.npz", {"wav": observed}),
+        "full_reference": ("full_reference_segment.npz", {"wav": full_segment}),
+    }
+    for name, (filename, arrays) in files.items():
+        np.savez_compressed(args.output_dir / filename, **arrays)
+    report = {
+        "scope": "one real generated-code 25-frame Code2Wav window on local CPU; no device execution",
+        "model_revision": args.model.resolve().name,
+        "model_weight_sha256": sha256(args.model / "model.safetensors"),
+        "decoder_weight_sha256": sha256(args.model / "speech_tokenizer" / "model.safetensors"),
+        "onnx_sha256": sha256(args.onnx),
+        "torch": torch.__version__,
+        "onnxruntime": ort.__version__,
+        "qwen_tts_sources": {
+            "wrapper": {"path": inspect.getfile(Qwen3TTSModel),
+                        "sha256": sha256(Path(inspect.getfile(Qwen3TTSModel)))},
+            "decoder": {"path": inspect.getfile(Qwen3TTSTokenizerV2Decoder),
+                        "sha256": sha256(Path(inspect.getfile(Qwen3TTSTokenizerV2Decoder)))},
+        },
+        "text": args.text,
+        "speaker": "Ryan",
+        "language": "English",
+        "seed": 42,
+        "generator_dtype": "bfloat16",
+        "decoder_dtype": "float32",
+        "threads": args.threads,
+        "generated_frames": int(codes.shape[0]),
+        "generated_audio_samples": len(generated),
+        "window": {"start_frame": 0, "history_frames": 72, "new_frames": 25},
+        "files": {name: {"filename": filename, "sha256": sha256(args.output_dir / filename)}
+                  for name, (filename, _) in files.items()},
+        "ort_vs_window_eager": compare(eager, observed),
+        "window_eager_vs_full_fp32_decoder_segment": compare(full_segment, eager),
+        "limits": [
+            "One generated stream and one window do not establish speech quality.",
+            "The model generates codes on local CPU; this does not test a mobile talker or handoff.",
+            "The FP32 full decoder reference is separate from the BF16 generator's waveform.",
+        ],
+    }
+    (args.output_dir / "report.json").write_text(json.dumps(report, indent=2) + "\n")
+    print(json.dumps({"frames": report["generated_frames"],
+                      "ort_vs_eager": report["ort_vs_window_eager"],
+                      "window_vs_full": report["window_eager_vs_full_fp32_decoder_segment"]}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/edge_harness/results/e2e_expansion_20260924/README.md

@@ -34,7 +34,14 @@ from that pinned snapshot matches local eager CPU at 1.99e-6 relative L2 and
 has compiled to S25 TFLite. Its GPU-requested inference returned an
 unsaturated waveform at 4.239% relative L2 versus local ONNX CPU, bitwise
 identical to the historical GPU output on this fixture. Its own placement
-profile remains pending; the historical timing is separate evidence.
+profile measured 100 component samples at p50/p95 **0.671/1.454 s** for
+2 s audio (median RTF **0.335**), with 403 CPU and 335 GPU rows. The
+historical profile is a separate run, and neither establishes full streaming.
+A [generated-code window](evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/README.md)
+from one pinned CustomVoice utterance passed local FP32 window/full-decoder
+continuity; the pinned S25 GPU-requested target returned an unsaturated
+waveform at **1.153% relative L2 / 38.76 dB SNR** versus local ONNX CPU.
+No listening-quality tolerance or complete device-local stream passed.
 
 The same pinned [two-frame graph on Galaxy S24](evidence/qwen_tts_vocoder_qualcomm/s24/tflite_short_chunk/README.md)
 matched local ONNX CPU at 5.75e-6 relative L2 when requested on CPU, while
@@ -45,8 +52,10 @@ A FP32-preserving GPU control returned the same saturated waveform bitwise.
 A
 [Snapdragon X Elite FP16 QNN short graph](evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/README.md)
 compiled and returned an unsaturated NPU-requested waveform at 0.930%
-relative L2 versus local ONNX CPU on one fixture. Its placement profile
-remains pending. These are component-only results.
+relative L2 versus local ONNX CPU on one fixture. Its 100-sample profile
+attributed all 728 rows to NPU, but p50/p95 was **3.662/3.676 s** for
+0.16 s audio (median component RTF **22.89**). These are component-only
+results and the short NPU vocoder is too slow for playable streaming.
 
 The [Qwen3-TTS vocoder study](evidence/qwen_tts_vocoder_qualcomm/README.md) reuses a retained fixed-shape Code2Wav export and historical Workbench input to test numerical behavior on SA8775P and other Qualcomm settings. The fresh SA8775P GPU-requested inference produces a saturated waveform, while the same binary's CPU-requested output on the exact SA device matches ONNX CPU within 4.13e-6 relative L2. Its 100-sample CPU-requested component profile had p50/p95 5.384/5.534 s with all 738 execution-detail rows on CPU. A FP32-preserving GPU-option control still saturated 84.89% of samples, so that option did not repair the tested path. RB3 GPU-requested execution produces an unsaturated but 15.73%-divergent waveform; its fresh 100-sample mixed CPU/GPU component p50/p95 was 3.099/3.133 s, not a usable TTS timing claim. On the exact RB3 device, the same binary requested on CPU matched original ONNX CPU within 4.13e-6 relative L2 on the retained fixture; its 47-sample component profile measured p50/p95 12.727/13.424 s with all 738 execution-detail rows on CPU. This is a major streaming-latency blocker. A single-fixture-calibrated RB3 INT8 QNN vocoder returned a finite but 66.93%-relative-L2 divergent waveform (3.49 dB SNR) versus ONNX CPU; its 100-sample component profile was p50/p95 1.971/1.992 s with all 728 execution-detail rows on NPU; these timings describe numerically unqualified audio. X Elite ONNX CPU output matches the local ONNX Runtime reference on the retained fixture; its 94-sample component p50/p95 was 6.384/6.644 s with all execution-detail rows on CPU. Corrected X Elite DirectML inference and profile both failed during provider initialization; separate FP16 QNN DLC inference returned an unsaturated waveform at 1.138% relative L2 / 38.88 dB SNR against original ONNX CPU; its placement profile timed out during QNN graph preparation without samples. A separate SA8775P FP16 QNN vocoder compile passed, but same-fixture NPU-requested inference failed after compilation without a waveform or placement samples. Fresh S24 and S25 CPU-requested TFLite vocoder inferences matched ONNX CPU within 4.04e-6 relative L2 on the retained fixture. S25 measured 100 component samples at p50/p95 4.418/4.641 s with all 738 execution-detail rows on CPU; S24 measured 100 at 4.826/4.958 s with all 738 rows on CPU. Separate explicit INT8 and W8A16 QDQ ONNX exports checked on local ORT CPU failed the same-fixture numerical gate at 246.97% and 331.83% waveform relative L2, respectively. S25 CPU talker inference preserved token 80 and its 100-sample component profile measured p50/p95 30.161/31.301 ms with all 1,800 rows on CPU. No complete on-device TTS stream or model-quality qualification is claimed.
 

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md

@@ -16,14 +16,23 @@ matched its eager decoder at 1.99e-6 relative L2 on the retained synthetic
 fixture and compiled for S25 TFLite. Its GPU-requested inference returned a
 finite unsaturated waveform at 4.239% relative L2 versus local ONNX CPU,
 bitwise identical to the historical target's device output on the same fixture.
-Its own placement profile is still running; the historical target's latency
-is separate evidence.
+Its own 100-sample placement profile measured p50/p95 **0.671/1.454 s** for
+2 s audio (median component RTF **0.335**), with 403 CPU and 335 GPU rows.
+This is a separate run from the historical target's 0.839/1.157 s profile;
+neither is a complete-stream or listening-quality result.
+A [generated-code window](s25/real_code_window/README.md) from one pinned
+CustomVoice utterance matched local FP32 ONNX/eager and full-decoder segment
+within 1.02e-6 and 7.26e-8 relative L2. The same S25 GPU-requested target
+returned an unsaturated waveform at **1.153% relative L2 / 38.76 dB SNR**
+versus ONNX CPU on that real-code window. This has no listening-quality or
+full-pipeline gate.
 
 The same pinned two-frame source [compiled for X Elite QNN](xelite/qnn_short_chunk/README.md)
 and its NPU-requested inference returned an unsaturated waveform at 0.930%
-relative L2 / 40.63 dB SNR versus local ONNX CPU. Its placement profile is
-still running; this is component numerical evidence, not verified NPU node
-placement or complete TTS.
+relative L2 / 40.63 dB SNR versus local ONNX CPU. Its 100-sample placement
+profile attributed all 728 rows to NPU, but p50/p95 was **3.662/3.676 s**
+for only 0.16 s audio (median component RTF **22.89**). This is a measured
+playable-stream blocker for that artifact, not complete TTS.
 
 On [Galaxy S24](s24/tflite_short_chunk/README.md), the same pinned two-frame TFLite source matched local ONNX CPU at 5.75e-6 relative L2 when requested on CPU, but its GPU-requested output saturated every sample and differed by 24.218 relative L2. The 100-sample CPU component p50/p95 was 3.433/3.873 s for only 0.16 s audio (median RTF 21.46), with all 738 rows on CPU. The mixed CPU/GPU component measured 0.601/0.895 s with 403 CPU and 335 GPU rows, but that timing describes numerically failed audio. An [FP32-preserving delegate control](s24/tflite_short_chunk_gpu_fp32/audit_report.json) returned the same saturated waveform bitwise, so that option did not fix this artifact. Neither route is a complete S24 TTS stream.
 

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_pinned_tflite/README.md

@@ -20,6 +20,18 @@ The new device waveform is [bitwise identical](historical_device_parity_report.j
 to the [historical 25-frame target's](../full_tflite_gpu_profile/README.md)
 GPU output on this retained fixture. This supports numerical continuity on
 one input, not identical historical export provenance or listening quality.
-A separate [placement profile](profile_submission.json) is in progress; GPU
-node placement and latency for this revision-pinned target remain unverified.
-Complete-stream handoff and sustained behavior are also unverified.
+A separate [placement profile](profile_report.json) returned 100 component
+samples at nearest-rank p50/p95 **0.671/1.454 s** for 2 s audio, with 403 CPU
+and 335 GPU execution-detail rows and 380,100,608 bytes reported peak memory.
+The median component RTF is **0.335**. Workbench generated profile input of
+the same shape, while the numerical inference used the retained fixture; the
+[audit](audit_report.json) binds both jobs to the exact target. Complete-stream
+handoff, listening quality and sustained behavior remain unverified.
+
+A separate [real generated-code window](../real_code_window/README.md) passed
+local ONNX/eager parity at 1.02e-6 relative L2 and matched the corresponding
+full FP32 decode segment at 7.26e-8 relative L2. The same fixture has a
+successful inference on this pinned S25 target: the unsaturated device
+waveform differed from local ONNX CPU by 1.153% relative L2 / 38.76 dB SNR.
+This one real-code window is encouraging component evidence, not a speech
+quality tolerance or complete mobile stream.

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_pinned_tflite/audit_report.json

@@ -1,6 +1,6 @@
 {
   "scope": "one 25-frame Qwen3-TTS vocoder component on hosted Samsung Galaxy S25; no complete stream",
-  "status": "component_inference_numeric_measured",
+  "status": "component_inference_numeric_and_placement_measured",
   "device": {
     "name": "Samsung Galaxy S25",
     "os": "15",
@@ -25,7 +25,8 @@
   "input_dataset_id": "d2nmpxd57",
   "jobs": {
     "compile": "jgddd1zzg",
-    "inference": "jglyymx25"
+    "inference": "jglyymx25",
+    "profile": "j568847ng"
   },
   "route": {
     "compile_options": "--target_runtime tflite",
@@ -72,7 +73,18 @@
   "limits": [
     "One fixed-window waveform has no listening-quality tolerance.",
     "The 72-frame input history, talker, predictor and stage handoff did not run on the device.",
-    "No persistent state, admission, cancellation or complete TTS stream was tested.",
-    "Requested accelerator execution is not verified node placement without a profile."
-  ]
+    "No persistent state, admission, cancellation or complete TTS stream was tested."
+  ],
+  "profile": {
+    "sample_count": 100,
+    "min_us": 560210,
+    "nearest_rank_p50_us": 670776,
+    "nearest_rank_p95_us": 1453828,
+    "reported_peak_memory_bytes": 380100608,
+    "compute_unit_row_counts": {
+      "CPU": 403,
+      "GPU": 335
+    },
+    "profile_input": "Workbench generated input with audited tensor shape; the waveform inference used the retained fixture"
+  }
 }

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/README.md

@@ -0,0 +1,27 @@
+# S25 pinned vocoder on generated speech codes
+
+The [reproducer](../../../../../../experiments/probe_qwen_tts_real_code_window.py)
+ran the pinned Qwen3-TTS 0.6B CustomVoice checkpoint at revision
+`85e237c12c027371202489a0ec509ded67b5e4b5` on HX370 WSL CPU. One
+seed-42 Ryan utterance yielded 117 frames of 16 codebook indices and a
+224,640-sample waveform. It decoded frames 0–96 as 72 history plus 25 new
+frames through the FP32 decoder and the revision-pinned ONNX export.
+
+The [local report](report.json) shows ONNX Runtime CPU versus the exact
+25-frame eager window at **1.02e-6 relative L2**. The eager window versus
+the same segment from an FP32 full 117-frame decode differed by **7.26e-8
+relative L2**. The [code tensor](codes.npz), [quantized fixture](fixture.npz),
+[eager](eager_output.npz), [ONNX CPU](ort_output.npz) and
+[full-decoder segment](full_reference_segment.npz) are retained. This checks
+the window boundary on one real generated-code stream; it is not a speech
+quality or mobile pipeline pass.
+
+The quantized fixture was uploaded as [Workbench dataset](dataset_submission.json)
+and [GPU-requested inference](inference_report.json) succeeded against the
+exact pinned S25 TFLite target. The [device waveform](device_output.npz) was
+finite and unsaturated; the [audit](audit_report.json) measures **1.153%
+relative L2 / 38.76 dB SNR** versus local ONNX CPU on this generated-code
+window. The separate 100-sample component profile used Workbench-generated
+input of the same shape and does not measure this specific fixture. One
+generated window does not establish speech intelligibility, all utterances,
+or device-local talker/predictor/vocoder integration.

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/audit_report.json

@@ -0,0 +1,85 @@
+{
+  "scope": "one generated-code 25-frame vocoder window on hosted S25; no device-local talker or complete stream",
+  "status": "component_numeric_measured_no_speech_quality_gate",
+  "device": {
+    "name": "Samsung Galaxy S25",
+    "os": "15",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8-elite-for-galaxy",
+      "chipset:sm8750-ac",
+      "hexagon:v79",
+      "soc-model:69",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "source_model_id": "mng2vkdrm",
+  "target_model_id": "mn4o5j17q",
+  "input_dataset_id": "d2q4y5jl7",
+  "jobs": {
+    "compile": "jgddd1zzg",
+    "inference": "j5688yryg"
+  },
+  "files": {
+    "fixture": {
+      "sha256": "4220eec4ee00e5fbb0df6f195d883c6d4549accf399853c3a8e2f83821724452",
+      "bytes": 185626
+    },
+    "device": {
+      "sha256": "7ecafcc289f098aa497f9c9eaa1da0e19879386157e8d2af7c1842e5ea1bfe46",
+      "bytes": 112200
+    },
+    "ort": {
+      "sha256": "df4ad896c62a0d621c9bc48120f929d3561666359a570068b01a3228354411a0",
+      "bytes": 179186
+    },
+    "eager": {
+      "sha256": "8c0c55197daa99433e8a7594525748b4e50e0570ef983dc5b76eedb56936d6d4",
+      "bytes": 179198
+    },
+    "full": {
+      "sha256": "3571033b52c4715761b469e17915d7b1456f5c251322c8a718fad8f905d428e9",
+      "bytes": 179207
+    }
+  },
+  "ort_cpu_vs_device": {
+    "relative_l2": 0.011533186351546298,
+    "snr_db": 38.76101381296397,
+    "max_abs": 0.011052757501602173,
+    "reference_rms": 0.074840782807901,
+    "device_rms": 0.07483006310024524,
+    "device_saturated_fraction_abs_ge_0_999": 0.0
+  },
+  "eager_cpu_vs_device": {
+    "relative_l2": 0.011533295679175895,
+    "snr_db": 38.760931476374985,
+    "max_abs": 0.011053577065467834,
+    "reference_rms": 0.0748407755479042,
+    "device_rms": 0.07483006310024524,
+    "device_saturated_fraction_abs_ge_0_999": 0.0
+  },
+  "full_fp32_decoder_segment_vs_device": {
+    "relative_l2": 0.011533293686009369,
+    "snr_db": 38.76093297745728,
+    "max_abs": 0.011053577065467834,
+    "reference_rms": 0.07484077531515666,
+    "device_rms": 0.07483006310024524,
+    "device_saturated_fraction_abs_ge_0_999": 0.0
+  },
+  "gross_numeric_failure_screen": {
+    "failed": false,
+    "criteria": "relative L2 >= 1.0 or >=10% saturated samples; not a speech-quality tolerance"
+  },
+  "limits": [
+    "One generated-code window is not a speech listening or intelligibility test.",
+    "Talker, predictor, co-resident handoff, state, cancellation and memory admission did not run on S25.",
+    "The separate placement profile used Workbench-generated same-shape input, not this fixture."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/dataset_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-24T06:46:09.898281+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "fixture_sha256": "4220eec4ee00e5fbb0df6f195d883c6d4549accf399853c3a8e2f83821724452",
+  "dataset_id": "d2q4y5jl7",
+  "input_shape": [
+    1,
+    512,
+    97
+  ],
+  "sample_count": 1,
+  "status": "UPLOADED"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/inference_report.json

@@ -0,0 +1,33 @@
+{
+  "timestamp_utc": "2026-09-24T06:53:00.381461+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5688yryg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5688yryg/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit gpu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "Samsung Galaxy S25",
+    "os": "15",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "abi:aarch64-android",
+      "vendor:samsung",
+      "format:phone",
+      "chipset:qualcomm-snapdragon-8-elite-for-galaxy",
+      "chipset:sm8750-ac",
+      "hexagon:v79",
+      "soc-model:69",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mn4o5j17q",
+  "input_dataset_id": "d2q4y5jl7",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/inference_submission.json

@@ -0,0 +1,12 @@
+{
+  "timestamp_utc": "2026-09-24T06:46:12.071383+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "j5688yryg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5688yryg/",
+  "model_id": "mn4o5j17q",
+  "input_dataset_id": "d2q4y5jl7",
+  "options": "--compute_unit gpu",
+  "device": "Samsung Galaxy S25",
+  "status": "SUBMITTED"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/report.json

@@ -0,0 +1,74 @@
+{
+  "scope": "one real generated-code 25-frame Code2Wav window on local CPU; no device execution",
+  "model_revision": "85e237c12c027371202489a0ec509ded67b5e4b5",
+  "model_weight_sha256": "bc3c7e785eb961179c25450d1acff03f839e0002f2f3a5aeb67b5735c0fa2adb",
+  "decoder_weight_sha256": "836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258",
+  "onnx_sha256": "01ee1cd02e5cd264200e5c50767a1805a6180378b051ff9b7ec0d8b6c75d7989",
+  "torch": "2.13.0+cpu",
+  "onnxruntime": "1.29.0",
+  "text": "The small red square is centered on the white card. I can see its straight edges clearly, and the surrounding background is bright and uncluttered.",
+  "speaker": "Ryan",
+  "language": "English",
+  "seed": 42,
+  "generator_dtype": "bfloat16",
+  "decoder_dtype": "float32",
+  "threads": 8,
+  "generated_frames": 117,
+  "generated_audio_samples": 224640,
+  "window": {
+    "start_frame": 0,
+    "history_frames": 72,
+    "new_frames": 25
+  },
+  "files": {
+    "codes": {
+      "filename": "codes.npz",
+      "sha256": "2768b9274748d767deca67ebd4c0201b914f4839902b89d5693d39e9b8e8c946"
+    },
+    "fixture": {
+      "filename": "fixture.npz",
+      "sha256": "4220eec4ee00e5fbb0df6f195d883c6d4549accf399853c3a8e2f83821724452"
+    },
+    "eager": {
+      "filename": "eager_output.npz",
+      "sha256": "8c0c55197daa99433e8a7594525748b4e50e0570ef983dc5b76eedb56936d6d4"
+    },
+    "ort": {
+      "filename": "ort_output.npz",
+      "sha256": "df4ad896c62a0d621c9bc48120f929d3561666359a570068b01a3228354411a0"
+    },
+    "full_reference": {
+      "filename": "full_reference_segment.npz",
+      "sha256": "3571033b52c4715761b469e17915d7b1456f5c251322c8a718fad8f905d428e9"
+    }
+  },
+  "ort_vs_window_eager": {
+    "relative_l2": 1.01682301490854e-06,
+    "snr_db": 119.85509264926462,
+    "max_abs": 1.1324882507324219e-06,
+    "reference_rms": 0.0748407755479042,
+    "observed_rms": 0.074840782807901
+  },
+  "window_eager_vs_full_fp32_decoder_segment": {
+    "relative_l2": 7.264658178733583e-08,
+    "snr_db": 142.7756963116475,
+    "max_abs": 1.695007085800171e-07,
+    "reference_rms": 0.07484077531515666,
+    "observed_rms": 0.0748407755479042
+  },
+  "limits": [
+    "One generated stream and one window do not establish speech quality.",
+    "The model generates codes on local CPU; this does not test a mobile talker or handoff.",
+    "The FP32 full decoder reference is separate from the BF16 generator's waveform."
+  ],
+  "qwen_tts_sources": {
+    "wrapper": {
+      "path": "/home/zhout/project/edge_infer/scratch/qwentts_wsl_pins/qwen_tts/inference/qwen3_tts_model.py",
+      "sha256": "e1da450732857c1f5fe3e36ebab85db2f6dc6a48caaff6973f463384e30275e4"
+    },
+    "decoder": {
+      "path": "/home/zhout/project/edge_infer/scratch/qwentts_wsl_pins/qwen_tts/core/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py",
+      "sha256": "844e8dd8c0182ef9c6463c874631c22ef3c5a4fd1899dd657016164cc5379628"
+    }
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/README.md

@@ -6,9 +6,14 @@ compiled to [FP16 QNN DLC](compile_report.json) for Snapdragon X Elite CRD.
 fixture. The [audit](audit_report.json) measures **0.930% relative L2** and
 **40.63 dB SNR** against local ONNX Runtime CPU, with no saturated samples.
 This is one synthetic, two-frame waveform without a listening-quality gate.
-A separate [placement profile](profile_submission.json) remains in progress
-for the same target and tensor shape; the request alone does not verify NPU
-node placement or latency. The earlier
+A separate [placement profile](profile_report.json) returned 100 component
+samples at nearest-rank p50/p95 **3.662/3.676 s** for **0.16 s** audio,
+with all 728 execution-detail rows on NPU and 15,855,616 bytes reported
+inference peak. Its median component RTF is **22.89**. The profile used
+Workbench-generated same-shape input, while the waveform inference used the
+retained fixture; the [audit](audit_report.json) binds both jobs to the exact
+target. This measured latency is a playable-stream blocker for this artifact.
+The earlier
 [25-frame X Elite NPU attempt](../qnn_npu/inference_report.json) returned a
 waveform but its placement profile timed out during graph preparation; this
 shorter artifact tests a separate memory/graph-preparation boundary. Neither

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/audit_report.json

@@ -1,6 +1,6 @@
 {
   "scope": "one 2-frame Qwen3-TTS vocoder component on hosted Snapdragon X Elite CRD; no complete stream",
-  "status": "component_inference_numeric_measured",
+  "status": "component_inference_numeric_and_placement_measured",
   "device": {
     "name": "Snapdragon X Elite CRD",
     "os": "11",
@@ -24,7 +24,8 @@
   "input_dataset_id": "d7ovqp0w2",
   "jobs": {
     "compile": "j5m0m22qg",
-    "inference": "jgdd3lxeg"
+    "inference": "jgdd3lxeg",
+    "profile": "jp4yyrrvp"
   },
   "route": {
     "compile_options": "--target_runtime qnn_dlc --qnn_options default_graph_htp_precision=FLOAT16",
@@ -71,7 +72,17 @@
   "limits": [
     "One fixed-window waveform has no listening-quality tolerance.",
     "The 72-frame input history, talker, predictor and stage handoff did not run on the device.",
-    "No persistent state, admission, cancellation or complete TTS stream was tested.",
-    "Requested accelerator execution is not verified node placement without a profile."
-  ]
+    "No persistent state, admission, cancellation or complete TTS stream was tested."
+  ],
+  "profile": {
+    "sample_count": 100,
+    "min_us": 3646540,
+    "nearest_rank_p50_us": 3661691,
+    "nearest_rank_p95_us": 3676170,
+    "reported_peak_memory_bytes": 15855616,
+    "compute_unit_row_counts": {
+      "NPU": 728
+    },
+    "profile_input": "Workbench generated input with audited tensor shape; the waveform inference used the retained fixture"
+  }
 }

benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md

@@ -49,8 +49,8 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | scoped E2E: Omni Radeon+CPU image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU+Radeon Cosmos (WSL/native Windows) |
 | PC HX370 + AMD NPU / Windows worker | NOT E2E (output-head component passed) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
-| PC Snapdragon X Elite CRD / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; two-frame QNN vocoder waveform measured, placement pending) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
-| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | NOT E2E (one Spark attention layer on NPU) | NOT E2E (FP16 talker token mismatch; CPU talker + vocoder profiled; 25-frame and two-frame QNN vocoder memory failures; historical 25-frame TFLite mixed vocoder component RTF 0.420; pinned 25-frame GPU waveform measured, placement pending; short chunk RTF 3.49) | NOT E2E | NOT E2E (NPU speech head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
+| PC Snapdragon X Elite CRD / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; two-frame QNN vocoder on NPU, component RTF 22.89) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
+| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | NOT E2E (one Spark attention layer on NPU) | NOT E2E (FP16 talker token mismatch; CPU talker + vocoder profiled; 25-frame and two-frame QNN vocoder memory failures; historical 25-frame TFLite mixed vocoder component RTF 0.420; pinned 25-frame mixed CPU/GPU vocoder profiled at RTF 0.335; generated-code waveform measured; short chunk RTF 3.49) | NOT E2E | NOT E2E (NPU speech head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | NOT E2E (one Spark attention layer on NPU) | NOT E2E (FP16 talker mismatch; short CPU vocoder RTF 21.46; mixed CPU/GPU waveform saturated) | NOT E2E | NOT E2E (NPU speech head step profiled) | NOT E2E (Cosmos encoder numeric and NPU profiled) |
 | Embedded SA8775P ADP / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; NPU vocoder failed) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | NOT E2E (Spark CPU layer passed; W8A8 NPU quality open) | NOT E2E (CPU talker + vocoder profiled; INT8 NPU vocoder fails numeric gate) | NOT E2E | NOT E2E (CPU speech-head step profiled) | NOT E2E (Cosmos CPU numeric and profiled) |
@@ -95,12 +95,12 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | InternVLA-A1 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC Snapdragon X Elite CRD / AI Hub | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | A real Spark-X2.5-1.7B one-layer W8A16 QDQ attention artifact, compiled as QNN DLC, ran on the exact hosted Snapdragon X Elite CRD (Windows 11 ARM64). Inference returned finite hidden/new-K/new-V tensors bitwise identical to the prior S24/S25 NPU outputs on the same fixture; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.433/1.606 ms; the 1.346 ms estimated-time field is the minimum. Hidden/K/V relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650%, without a next-token quality tolerance. There is no complete 28-layer prefill/decode/KV-ring/sampling loop, resident Omni controller or complete-request timing. | Qualify a complete X Elite resident artifact/controller with state, token quality, memory admission/cancel and sustained power; [X Elite NPU raw evidence](../../../e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md). |
-| PC Snapdragon X Elite CRD / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_ONNX_CPU_VOCODER_PARITY_PROFILED; DIRECTML_RUNTIME_FAILED; TALKER_FP16_NUMERIC_FAIL; QNN_NPU_VOCODER_NUMERIC_MEASURED_PROFILE_TIMEOUT; SHORT_QNN_VOCODER_NUMERIC_MEASURED_PLACEMENT_PENDING; TALKER_ONNX_CPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The fresh ONNX vocoder inference on the exact Windows 11 ARM64 X Elite CRD requested CPU and returned an unsaturated waveform on the pinned fixture, matching local ONNX Runtime CPU with 8.09e-7 relative L2 and 121.84 dB SNR. Its 94-sample component profile had nearest-rank p50/p95 6.384/6.644 s and all 634 execution-detail rows on CPU; 5.917 s estimated time is the sample minimum. This is one component, without a listening-quality gate. Corrected DirectML inference and profile both failed at provider initialization with Windows error 80070057. A separate FP16 QNN DLC vocoder inference returned an unsaturated waveform at 1.138% relative L2 / 38.88 dB SNR versus local ONNX CPU on the same fixture; the NPU node-placement profile timed out during graph preparation with no timing or per-node samples, and no listening-quality tolerance is set. A new [two-frame QNN vocoder candidate](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/README.md) compiled for X Elite and returned an unsaturated waveform at 0.930% relative L2 / 40.63 dB SNR versus pinned local ONNX CPU on one synthetic fixture. Its NPU-requested placement profile remains in progress; inference request alone does not verify NPU nodes or latency. The 28-layer FP16 QNN talker step executed, yet changed CPU top-1 token 80 to 1720 with 8.57% logits relative L2 on a synthetic cache fixture; a separate ONNX CPU-requested talker step preserved CPU token 80 with logits relative L2 2.07e-5 and maximum K/V relative L2 2.18e-5. Its 100-sample component profile had p50/p95 191.414/309.428 ms and all 1,744 execution-detail rows on CPU. The older TFLite vocoder target was rejected on Windows ARM; no complete stream exists. | Complete the short QNN vocoder placement profile and audit CPU talker placement; investigate DirectML failure, then qualify complete local streaming, memory and sustained behavior; [vocoder](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md) and [talker](../../../e2e_expansion_20260924/evidence/qwen_tts_talker_qualcomm/README.md) evidence. |
+| PC Snapdragon X Elite CRD / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_ONNX_CPU_VOCODER_PARITY_PROFILED; DIRECTML_RUNTIME_FAILED; TALKER_FP16_NUMERIC_FAIL; QNN_NPU_VOCODER_NUMERIC_MEASURED_PROFILE_TIMEOUT; SHORT_QNN_NPU_VOCODER_NUMERIC_PROFILED_SLOW; TALKER_ONNX_CPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The fresh ONNX vocoder inference on the exact Windows 11 ARM64 X Elite CRD requested CPU and returned an unsaturated waveform on the pinned fixture, matching local ONNX Runtime CPU with 8.09e-7 relative L2 and 121.84 dB SNR. Its 94-sample component profile had nearest-rank p50/p95 6.384/6.644 s and all 634 execution-detail rows on CPU; 5.917 s estimated time is the sample minimum. This is one component, without a listening-quality gate. Corrected DirectML inference and profile both failed at provider initialization with Windows error 80070057. A separate FP16 QNN DLC vocoder inference returned an unsaturated waveform at 1.138% relative L2 / 38.88 dB SNR versus local ONNX CPU on the same fixture; the NPU node-placement profile timed out during graph preparation with no timing or per-node samples, and no listening-quality tolerance is set. A new [two-frame QNN vocoder candidate](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/README.md) compiled for X Elite and returned an unsaturated waveform at 0.930% relative L2 / 40.63 dB SNR versus pinned local ONNX CPU on one synthetic fixture. Its 100-sample placement profile measured nearest-rank p50/p95 3.662/3.676 s for 0.16 s audio, all 728 execution rows on NPU, 15,855,616 bytes reported peak and median component RTF 22.89. The profile used Workbench-generated same-shape input; the waveform used the retained fixture. This is too slow for playable streaming on the tested artifact. The 28-layer FP16 QNN talker step executed, yet changed CPU top-1 token 80 to 1720 with 8.57% logits relative L2 on a synthetic cache fixture; a separate ONNX CPU-requested talker step preserved CPU token 80 with logits relative L2 2.07e-5 and maximum K/V relative L2 2.18e-5. Its 100-sample component profile had p50/p95 191.414/309.428 ms and all 1,744 execution-detail rows on CPU. The older TFLite vocoder target was rejected on Windows ARM; no complete stream exists. | Find a quality-passing, much faster vocoder path and audit CPU talker placement; investigate DirectML failure, then qualify complete local streaming, memory and sustained behavior; [vocoder](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md) and [talker](../../../e2e_expansion_20260924/evidence/qwen_tts_talker_qualcomm/README.md) evidence. |
 | PC Snapdragon X Elite CRD / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC Snapdragon X Elite CRD / AI Hub | MiniCPM-o 4.5 | COMPONENT_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real 20-layer MiniCPM-o speech-head FP16 QNN DLC ran on the exact X Elite CRD. Same-fixture inference returned 41 finite tensors, preserved ONNX CPU top-1 speech token 1867, and had 0.3170% logits relative L2 and 1.221% maximum K/V relative L2. Its 100-sample component profile had nearest-rank p50/p95 9.905/11.037 ms, with all 944 execution-detail rows on NPU. Reported inference peak memory was 79,208,448 bytes and first-load peak 1,081,573,376 bytes. Thinker/encoders/vocoder and a device-local stateful Omni backend remain unqualified. | Validate the complete multimodal pipeline, state, quality, co-resident memory and sustained behavior; [X Elite raw evidence](../../../e2e_expansion_20260924/evidence/minicpmo_tts_qualcomm/README.md). |
 | PC Snapdragon X Elite CRD / AI Hub | InternVLA-A1 | COMPONENT_COSMOS_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos FP16 QNN image encoder returned a finite FP32 latent for the six-frame synthetic pattern on exact X Elite CRD. Against the pinned source, relative L2 was 0.0608%, SNR 64.32 dB and cosine 0.999999815. The 100-sample six-frame encoder profile had p50/p95 82.939/83.837 ms, with all 121 execution-detail rows on NPU. Reported component inference peak memory was 24,059,904 bytes and first-load peak 692,891,648 bytes. No real-observation action tolerance, complete policy or local stateful Omni backend exists. | Validate real observations, action units/timing, state, admission, cancellation and sustained behavior; [X Elite Cosmos raw evidence](../../../e2e_expansion_20260924/evidence/internvla_cosmos_qualcomm/README.md). |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | The real 1.7B checkpoint one-layer W8A16 full-attention decoder artifact executed for a fixed 1,024-token cache bucket. Same-fixture outputs matched S24; the hosted S25 profile had 100 component samples at nearest-rank p50/p95 0.957/1.028 ms with all 73 execution-detail rows on NPU. FP32-source hidden-state relative L2 was 1.351% without a downstream token-quality gate. No full 28-layer prefill/decode, KV/ring, sampling or resident controller ran on S25. | Implement and validate local 128-token generation, 512/1024 transitions, cancellation, resident memory and sustained power; [S25/S24 component evidence](../../../e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md). |
-| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3-TTS 0.6B CustomVoice | TALKER_FP16_NUMERIC_FAIL; CPU_TALKER_AND_VOCODER_NUMERIC_PROFILED; QNN_VOCODER_MEMORY_FAIL; FULL_TFLITE_MIXED_COMPONENT_PROFILED; PINNED_FULL_TFLITE_GPU_NUMERIC_MEASURED_PLACEMENT_PENDING; SHORT_TFLITE_MIXED_COMPONENT_PROFILED; NOT_E2E_PROFILEABLE | Fresh FP16 predictor inference passed; its one output exactly matches the same compiled artifact/dataset reference. The same historical S25 TFLite vocoder target requested on CPU returned a finite unsaturated waveform at 4.04e-6 relative L2 / 107.86 dB SNR versus original ONNX CPU on the retained fixture. Its 100-sample component profile measured nearest-rank p50/p95 4.418/4.641 s, with all 738 execution-detail rows on CPU and reported inference peak memory 867,893,248 bytes. A separate 28-layer TFLite CPU-requested talker step returned 57 finite outputs, preserved ONNX CPU top-1 speech token 80, and had logits relative L2 3.33e-5 and maximum new-cache relative L2 3.41e-5. Its 100-sample component profile measured p50/p95 30.161/31.301 ms, with all 1,800 execution-detail rows on CPU and reported inference peak memory 1,929,318,400 bytes. The FP16 QNN talker step executed on the synthetic cache fixture but changed CPU top-1 speech token 80 to 412 with 8.99% logits relative L2 and maximum new-K/V relative L2 11.71%. A separately compiled FP16 QNN DLC talker also changed top-1 to 1720 with 8.77% logits relative L2 and 11.27% maximum new-cache relative L2. Both tested NPU talker artifacts are numerically unqualified. A fresh 25-frame FP16 QNN DLC vocoder compiled for S25, but its NPU-requested inference failed when Workbench reported device memory limits; it returned no waveform or profile samples. A [two-frame short-window export](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/qnn_short_chunk/README.md) matched local ONNX CPU at 2.96e-6 relative L2, then its separately compiled S25 FP16 QNN DLC also failed NPU-requested inference with the same reported device-memory limit. Its local HX370 WSL CPU component p50/p95 was 1.365/1.453 s to produce 0.16 s audio, not an S25 estimate. A separate [two-frame TFLite GPU-requested inference](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/tflite_short_chunk/README.md) returned finite unsaturated audio at 2.325% relative L2 / 32.67 dB SNR versus local ONNX CPU on one fixture; its 100-sample placement profile measured p50/p95 0.559/1.052 s for 0.16 s audio, with 403 CPU and 335 GPU execution-detail rows and 379,121,664 bytes reported peak. The mixed CPU/GPU short component median RTF is 3.49. A separate [25-frame TFLite GPU-requested profile](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_tflite_gpu_profile/README.md) of the historical target measured 100 component samples at p50/p95 0.839/1.157 s for 2 s audio, with 403 CPU and 335 GPU rows, reported 374,591,488-byte peak and median component RTF 0.420. Its retained GPU waveform differed from ONNX CPU by 4.24% relative L2 on one fixture. A pinned CustomVoice eager 25-frame decoder matched the historical ONNX CPU waveform at 1.89e-6 relative L2 on that fixture, but the historical export revision and listening quality remain unverified. A fresh [revision-pinned 25-frame TFLite export](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_pinned_tflite/README.md) matched local eager CPU at 1.99e-6 relative L2 and compiled for S25. Its GPU-requested inference returned an unsaturated waveform at 4.239% relative L2 / 27.46 dB SNR versus pinned local ONNX CPU, bitwise identical to the historical GPU output on this one fixture. Its own placement profile remains pending; the historical target's timing does not transfer automatically. No complete co-resident NPU/GPU stream exists. | Improve and revalidate the [talker artifact](../../../e2e_expansion_20260924/evidence/qwen_tts_talker_qualcomm/README.md), then qualify the [device-local predictor/vocoder pipeline](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md), output quality, recovery and 30-minute thermal behavior. |
+| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3-TTS 0.6B CustomVoice | TALKER_FP16_NUMERIC_FAIL; CPU_TALKER_AND_VOCODER_NUMERIC_PROFILED; QNN_VOCODER_MEMORY_FAIL; FULL_TFLITE_MIXED_COMPONENT_PROFILED; PINNED_FULL_TFLITE_MIXED_COMPONENT_PROFILED; SHORT_TFLITE_MIXED_COMPONENT_PROFILED; NOT_E2E_PROFILEABLE | Fresh FP16 predictor inference passed; its one output exactly matches the same compiled artifact/dataset reference. The same historical S25 TFLite vocoder target requested on CPU returned a finite unsaturated waveform at 4.04e-6 relative L2 / 107.86 dB SNR versus original ONNX CPU on the retained fixture. Its 100-sample component profile measured nearest-rank p50/p95 4.418/4.641 s, with all 738 execution-detail rows on CPU and reported inference peak memory 867,893,248 bytes. A separate 28-layer TFLite CPU-requested talker step returned 57 finite outputs, preserved ONNX CPU top-1 speech token 80, and had logits relative L2 3.33e-5 and maximum new-cache relative L2 3.41e-5. Its 100-sample component profile measured p50/p95 30.161/31.301 ms, with all 1,800 execution-detail rows on CPU and reported inference peak memory 1,929,318,400 bytes. The FP16 QNN talker step executed on the synthetic cache fixture but changed CPU top-1 speech token 80 to 412 with 8.99% logits relative L2 and maximum new-K/V relative L2 11.71%. A separately compiled FP16 QNN DLC talker also changed top-1 to 1720 with 8.77% logits relative L2 and 11.27% maximum new-cache relative L2. Both tested NPU talker artifacts are numerically unqualified. A fresh 25-frame FP16 QNN DLC vocoder compiled for S25, but its NPU-requested inference failed when Workbench reported device memory limits; it returned no waveform or profile samples. A [two-frame short-window export](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/qnn_short_chunk/README.md) matched local ONNX CPU at 2.96e-6 relative L2, then its separately compiled S25 FP16 QNN DLC also failed NPU-requested inference with the same reported device-memory limit. Its local HX370 WSL CPU component p50/p95 was 1.365/1.453 s to produce 0.16 s audio, not an S25 estimate. A separate [two-frame TFLite GPU-requested inference](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/tflite_short_chunk/README.md) returned finite unsaturated audio at 2.325% relative L2 / 32.67 dB SNR versus local ONNX CPU on one fixture; its 100-sample placement profile measured p50/p95 0.559/1.052 s for 0.16 s audio, with 403 CPU and 335 GPU execution-detail rows and 379,121,664 bytes reported peak. The mixed CPU/GPU short component median RTF is 3.49. A separate [25-frame TFLite GPU-requested profile](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_tflite_gpu_profile/README.md) of the historical target measured 100 component samples at p50/p95 0.839/1.157 s for 2 s audio, with 403 CPU and 335 GPU rows, reported 374,591,488-byte peak and median component RTF 0.420. Its retained GPU waveform differed from ONNX CPU by 4.24% relative L2 on one fixture. A pinned CustomVoice eager 25-frame decoder matched the historical ONNX CPU waveform at 1.89e-6 relative L2 on that fixture, but the historical export revision and listening quality remain unverified. A fresh [revision-pinned 25-frame TFLite export](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/full_pinned_tflite/README.md) matched local eager CPU at 1.99e-6 relative L2 and compiled for S25. Its GPU-requested inference returned an unsaturated waveform at 4.239% relative L2 / 27.46 dB SNR versus pinned local ONNX CPU, bitwise identical to the historical GPU output on this one fixture. Its own 100-sample placement profile measured nearest-rank p50/p95 0.671/1.454 s for 2 s audio, median component RTF 0.335, with 403 CPU and 335 GPU execution rows and 380,100,608 bytes reported peak memory. The profile used Workbench-generated same-shape input; the numerical inference used the retained fixture. The historical 0.839/1.157 s profile is a separate run. A [real generated-code 25-frame window](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/README.md) passed local ONNX/eager parity at 1.02e-6 relative L2 and matched the corresponding full FP32 decoder segment at 7.26e-8 relative L2; the same pinned S25 GPU-requested target returned an unsaturated waveform at 1.153% relative L2 / 38.76 dB SNR versus local ONNX CPU on this generated-code window, without a listening-quality gate. No complete co-resident NPU/GPU stream exists. | Improve and revalidate the [talker artifact](../../../e2e_expansion_20260924/evidence/qwen_tts_talker_qualcomm/README.md), then qualify the [device-local predictor/vocoder pipeline](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md), output quality, recovery and 30-minute thermal behavior. |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | MiniCPM-o 4.5 | COMPONENT_FP16_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | Fresh FP16 speech-head inference passed; all 41 outputs exactly matched the same artifact/dataset reference. Against the recovered original ONNX CPU source on the same fixture, logits relative L2 was 0.3605%, maximum new-K/V relative L2 1.311%, and top-1 speech token 1867 matched. Historical FP16 parity passed; tested W8A16 top-1 parity failed. The 100-sample component profile had p50/p95 7.197/7.615 ms, with all 1,029 execution-detail rows on NPU and reported inference peak memory 185,991,168 bytes. Generated-audio quality and qualified thinker/encoders/vocoder remain unverified. | Qualify a complete mobile checkpoint/export and the full multimodal/stateful pipeline; [speech-head evidence](../../../e2e_expansion_20260924/evidence/minicpmo_tts_qualcomm/README.md). |
 | Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | InternVLA-A1 | COMPONENT_COSMOS_NPU_NUMERIC_PROFILED; SYNTHETIC_ACTION_SENSITIVITY_MEASURED; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos image encoder FP32 ONNX export matched its source on two six-frame fixtures. Galaxy S25 FP16 QNN DLC inference returned a finite latent within 0.0524% relative L2 / 65.61 dB SNR of the pinned pattern source. Injecting this device latent into the same real policy on local CPU changed synthetic `[1,50,32]` actions by 0.742% relative L2; the CPU source repeated bitwise identically. The 100-sample six-frame encoder profile had nearest-rank p50/p95 49.665/51.382 ms, all 121 execution-detail rows on NPU, and a reported inference peak of 158,232,576 bytes. No real-observation task tolerance exists, and the action policy was not resident on S25. ADB has no attached device. | Extend the [S25 encoder component](../../../e2e_expansion_20260924/evidence/internvla_cosmos_qualcomm/README.md), then validate real action quality, device-local state, admission, cancellation and sustained behavior. |

docs/design/omni_edge_e2e_roadmap.md

@@ -48,8 +48,16 @@ now matches local eager CPU at 1.99e-6 relative L2 and has compiled for S25
 TFLite. Its GPU-requested inference returned an unsaturated waveform at
 4.239% relative L2 versus pinned local ONNX CPU, bitwise identical to the
 historical target's output on the same fixture. The new target's own
-node-placement profile is pending, so the historical timing does not yet
-qualify this revision-pinned artifact.
+100-sample node-placement profile measured p50/p95 0.671/1.454 s for 2 s
+audio, with 403 CPU and 335 GPU rows (median component RTF 0.335). This
+qualifies only the tested vocoder component; full co-resident streaming,
+speech quality and sustained behavior remain open.
+A [real generated-code 25-frame window](../../benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25/real_code_window/README.md)
+also passed local ONNX/eager parity at 1.02e-6 relative L2 and matched its
+full FP32 decoder segment at 7.26e-8 relative L2. The corresponding S25
+GPU-requested inference returned an unsaturated waveform at 1.153% relative
+L2 / 38.76 dB SNR versus ONNX CPU. This tests one generated-code window,
+not mobile talker/vocoder co-residency or speech quality.
 On [Galaxy S24](../../benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s24/tflite_short_chunk/README.md),
 the same pinned two-frame TFLite source passed CPU-requested waveform parity
 at 5.75e-6 relative L2, but the GPU-requested waveform saturated completely
@@ -59,8 +67,10 @@ measured 0.601/0.895 s but failed the waveform gate. A FP32-preserving GPU
 option returned the same saturated waveform bitwise, so it did not repair
 the tested delegate artifact. A [Snapdragon X Elite short QNN graph](../../benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/qnn_short_chunk/README.md)
 compiled and returned an unsaturated NPU-requested waveform at 0.930%
-relative L2 versus local ONNX CPU on one synthetic fixture. Its node-placement
-profile is pending. Neither changes M2 support
+relative L2 versus local ONNX CPU on one synthetic fixture. Its 100-sample
+profile attributed all 728 rows to NPU, but p50/p95 was 3.662/3.676 s for
+0.16 s audio (median component RTF 22.89). This artifact therefore fails
+the playable-stream latency gate. Neither changes M2 support
 depth before full stream and quality gates.
 
 The [Spark hosted Qualcomm component expansion](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md) advances M1 only to one real fixed-shape decoder attention layer. The same W8A16 QNN DLC ran on S24, Snapdragon X Elite CRD and SA8775P ADP NPUs with identical outputs on one pinned fixture; S25 had prior separate component evidence. On [RB3 Gen 2](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md), the tested W8A16 graph failed QNN loading even after an exact-device compile. FP32 ONNX ran on CPU with near-source parity, while calibrated W8A8 ran on NPU but had 20.9% hidden-state relative L2 and remains numerically unqualified. These are component C results, not full M1: device-local 28-layer prefill/continuous decode, KV/ring, sampling, token quality, admission/cancellation, complete-request timing and sustained power/thermal gates remain open. AI Hub is the test facility, not a deployment dependency.