Audit Qualcomm Qwen3-TTS vocoder parity and profile

benchmarks/edge_harness/experiments/audit_qwen_tts_vocoder_qualcomm.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+"""Audit a hosted Qwen3-TTS vocoder component against a pinned ONNX CPU run."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+from collections import Counter
+from pathlib import Path
+
+import numpy as np
+
+
+SOURCE_SHA256 = "7c07666229c6d404894132e22f784525e3212a49efad55ddd532eda5286c2b2b"
+FIXTURE_SHA256 = "dfca9e8d2a724a568a31b6cd18ff6a52fdc205fd34e13cbf1cd75f0f7937e82d"
+CPU_OUTPUT_SHA256 = "8de78b94b59f01e49852f7706adbd2508a9976b8e2705cfd3b94d07bb3606cb1"
+
+
+def sha256(path: Path) -> str:
+    with path.open("rb") as source:
+        return hashlib.file_digest(source, "sha256").hexdigest()
+
+
+def read_output(path: Path, name: str) -> np.ndarray:
+    with np.load(path, allow_pickle=False) as data:
+        if data.files != [name]:
+            raise ValueError(f"{path}: expected only {name}, got {data.files}")
+        output = np.asarray(data[name])
+    if output.shape != (1, 48000) or output.dtype != np.float32 or not np.isfinite(output).all():
+        raise ValueError(f"{path}: violates finite FP32 48,000-sample WAV contract")
+    return output
+
+
+def compare(reference: np.ndarray, output: np.ndarray) -> dict:
+    a = reference.astype(np.float64).ravel()
+    b = output.astype(np.float64).ravel()
+    error_norm = float(np.linalg.norm(a - b))
+    reference_norm = float(np.linalg.norm(a))
+    output_norm = float(np.linalg.norm(b))
+    if reference_norm == 0 or output_norm == 0:
+        raise ValueError("silent reference/output cannot define waveform similarity")
+    return {
+        "relative_l2": error_norm / reference_norm,
+        "snr_db": 20 * math.log10(reference_norm / error_norm) if error_norm else None,
+        "max_abs": float(np.max(np.abs(a - b))),
+        "cosine": float(np.dot(a, b) / (reference_norm * output_norm)),
+        "reference_rms": float(np.sqrt(np.mean(a * a))),
+        "output_rms": float(np.sqrt(np.mean(b * b))),
+        "output_saturated_fraction_abs_ge_0_999": float(np.mean(np.abs(b) >= 0.999)),
+    }
+
+
+def nearest_rank(values: list[int], fraction: float) -> int:
+    ordered = sorted(values)
+    return ordered[math.ceil(fraction * len(ordered)) - 1]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--device-name", required=True)
+    parser.add_argument("--compile-options", required=True)
+    parser.add_argument("--run-options", required=True)
+    for name in (
+        "source-onnx", "fixture", "cpu-output", "s25-output", "s25-inference-report", "compile-report",
+        "inference-report", "profile-report", "device-output", "output-report",
+    ):
+        parser.add_argument(f"--{name}", type=Path, required=True)
+    args = parser.parse_args()
+    compile_report = json.loads(args.compile_report.read_text(encoding="utf-8"))
+    inference = json.loads(args.inference_report.read_text(encoding="utf-8"))
+    profile = json.loads(args.profile_report.read_text(encoding="utf-8"))
+    s25_inference = json.loads(args.s25_inference_report.read_text(encoding="utf-8"))
+    if any(report.get("status") != "SUCCESS" for report in (compile_report, inference, profile)):
+        raise RuntimeError("compile, inference and profile must all succeed")
+    if compile_report.get("source_model_id") != "mqyer339n" or compile_report.get("options") != args.compile_options:
+        raise RuntimeError("compile did not use pinned Qwen3-TTS vocoder source/options")
+    target_id = compile_report["target_model_id"]
+    if inference.get("model_id") != target_id or profile.get("model_id") != target_id:
+        raise RuntimeError("inference/profile used a different target artifact")
+    if inference.get("input_dataset_id") != "d7m80jl32":
+        raise RuntimeError("inference did not use the retained vocoder fixture")
+    if (s25_inference.get("status") != "SUCCESS" or s25_inference.get("model_id") != "mng7kwxon"
+            or s25_inference.get("input_dataset_id") != inference["input_dataset_id"]
+            or s25_inference.get("device", {}).get("name") != "Samsung Galaxy S25"):
+        raise RuntimeError("historical S25 output lacks matching model/device/dataset provenance")
+    for report in (compile_report, inference, profile):
+        if report.get("device", {}).get("name") != args.device_name:
+            raise RuntimeError("job did not use the exact requested device")
+    if inference.get("options") != args.run_options or profile.get("options") != args.run_options:
+        raise RuntimeError("inference/profile requested different compute options")
+    files = {
+        name: {"bytes": path.stat().st_size, "sha256": sha256(path)}
+        for name, path in {
+            "source_onnx": args.source_onnx,
+            "fixture": args.fixture,
+            "cpu_output": args.cpu_output,
+            "s25_output": args.s25_output,
+            "device_output": args.device_output,
+        }.items()
+    }
+    for name, expected in {
+        "source_onnx": SOURCE_SHA256,
+        "fixture": FIXTURE_SHA256,
+        "cpu_output": CPU_OUTPUT_SHA256,
+    }.items():
+        if files[name]["sha256"] != expected:
+            raise RuntimeError(f"{name} differs from the pinned source/fixture/reference")
+    with np.load(args.fixture, allow_pickle=False) as data:
+        if data.files != ["quantized"] or data["quantized"].shape != (1, 512, 97):
+            raise RuntimeError("fixture input contract differs")
+    cpu = read_output(args.cpu_output, "wav")
+    s25 = read_output(args.s25_output, "output_0__0")
+    output = read_output(args.device_output, "output_0__0")
+    execution = profile["profile"]
+    times_us = execution["execution_summary"]["all_inference_times"]
+    if not times_us or any(not isinstance(value, int) or value <= 0 for value in times_us):
+        raise RuntimeError("missing or invalid device samples")
+    units = Counter(row.get("compute_unit") for row in execution["execution_detail"])
+    if not units:
+        raise RuntimeError("no compute-unit placement detail")
+    report = {
+        "scope": "one Qwen3-TTS code2wav 48,000-sample component on one retained fixture; no complete TTS stream",
+        "status": "component_executed_quality_unqualified",
+        "device": profile["device"],
+        "compile_job_id": compile_report["job_id"],
+        "inference_job_id": inference["job_id"],
+        "profile_job_id": profile["job_id"],
+        "target_model_id": target_id,
+        "input_dataset_id": inference["input_dataset_id"],
+        "s25_prior_inference_job_id": s25_inference["job_id"],
+        "requested_run_options": args.run_options,
+        "files": files,
+        "cpu_source_vs_device": compare(cpu, output),
+        "s25_gpu_vs_device": compare(s25, output),
+        "profile": {
+            "sample_count": len(times_us),
+            "sample_min_us": min(times_us),
+            "nearest_rank_p50_us": nearest_rank(times_us, 0.50),
+            "nearest_rank_p95_us": nearest_rank(times_us, 0.95),
+            "reported_estimated_inference_time_us": execution["execution_summary"]["estimated_inference_time"],
+            "reported_peak_memory_bytes": execution["execution_summary"]["estimated_inference_peak_memory"],
+            "compute_unit_row_counts": dict(units),
+        },
+        "limits": [
+            "Historical source export does not attest its exact checkpoint revision.",
+            "This fixed component waveform has no listening or task-quality tolerance.",
+            "Hosted component profiling excludes talker, predictor, stage handoff, streaming, admission and sustained power/thermal behavior.",
+        ],
+    }
+    args.output_report.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    print(json.dumps({"status": report["status"], "profile": report["profile"]}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/edge_harness/experiments/collect_aihub_component.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Preserve raw Workbench Spark-layer jobs submitted in an evidence directory.
+"""Preserve raw Workbench component jobs submitted in an evidence directory.
 
 Requires qai-hub authentication through the local client configuration. The
 submission JSON files identify existing jobs; this script never resubmits them.
@@ -70,10 +70,11 @@ def collect(evidence_dir: Path, kind: str, client: hub.Client) -> str:
 def main() -> None:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("evidence_dirs", nargs="+", type=Path)
+    parser.add_argument("--kinds", nargs="+", choices=("inference", "profile"), default=("inference", "profile"))
     args = parser.parse_args()
     client = hub.Client()
     for directory in args.evidence_dirs:
-        for kind in ("inference", "profile"):
+        for kind in args.kinds:
             print(f"{directory.name}: {collect(directory, kind, client)}")
 
 

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md

@@ -11,7 +11,7 @@ Four routes were tested on exact device **Dragonwing RB3 Gen 2 Vision Kit**, usi
 | Compile original FP32 source to ONNX `mmxj02ljq`, request CPU | [Compile](compile_cpu_report.json) `jglyxowj5`, [inference](cpu/inference_report.json) `jgolrodqg` and [profile](cpu/profile_report.json) `jpvldemk5` passed. All 56 execution-detail rows were CPU. | [Output](cpu/device_output.npz) hidden/K/V relative L2 versus original ONNX Runtime 1.29 CPU was 8.73e-7/2.98e-7/3.37e-7. [Audit](cpu/audit_report.json): 100 samples, nearest-rank p50/p95 19.304/21.113 ms (minimum 19.104 ms). This is explicitly CPU, not NPU fallback masquerading as acceleration. |
 | Compile original source with calibration dataset `d7x8e3d59` using `--quantize_full_type int8 --quantize_io` to QNN DLC `mqkp10ezm`, request NPU | [Compile](compile_w8a8_report.json) `jpyoxq885`, [inference](w8a8/inference_report.json) `jgzlm8yo5` and [profile](w8a8/profile_report.json) `j5wl78z3p` passed. All 61 execution-detail rows were NPU. | [Output](w8a8/device_output.npz) hidden/K/V relative L2 versus FP32 CPU was **20.914%/1.702%/4.413%**. [Audit](w8a8/audit_report.json): 100 samples, p50/p95 5.610/5.761 ms (minimum 5.089 ms). This precision is **numerically unqualified**; no next-token/task tolerance is established. |
 
-All successful jobs used the same fixture dataset `d7m83gjy2` and returned finite FP32 hidden/new-key/new-value tensors. The service `estimated_inference_time` equals each sample minimum, not the median. The [CPU](cpu/audit_report.json) and [W8A8 NPU](w8a8/audit_report.json) audits pin source and fixture hashes, exact job/device/model IDs, output hashes, placement rows and full precision differences. [Submission records](compile_w8a8_submission.json), [initial failed submissions](inference_submission.json), [RB3 W8A16 compile submission](compile_rb3_submission.json), and [CPU compile submission](compile_cpu_submission.json) preserve attempted settings. The scripts [collect existing jobs](../../../../experiments/collect_spark_aihub_component.py) and [audit fallback variants](../../../../experiments/audit_spark_qualcomm_fallback.py) without embedded credentials.
+All successful jobs used the same fixture dataset `d7m83gjy2` and returned finite FP32 hidden/new-key/new-value tensors. The service `estimated_inference_time` equals each sample minimum, not the median. The [CPU](cpu/audit_report.json) and [W8A8 NPU](w8a8/audit_report.json) audits pin source and fixture hashes, exact job/device/model IDs, output hashes, placement rows and full precision differences. [Submission records](compile_w8a8_submission.json), [initial failed submissions](inference_submission.json), [RB3 W8A16 compile submission](compile_rb3_submission.json), and [CPU compile submission](compile_cpu_submission.json) preserve attempted settings. The scripts [collect existing jobs](../../../../experiments/collect_aihub_component.py) and [audit fallback variants](../../../../experiments/audit_spark_qualcomm_fallback.py) without embedded credentials.
 
 The CPU and NPU numbers are different precisions and outputs, not a qualified speedup comparison. Neither job includes 27 other decoder layers, prefill, continuous decode, KV/ring, sampling, a resident Omni controller, complete-request timing or sustained power/thermal behavior. Workbench is a test facility, not a deployment dependency.
 

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md

@@ -8,6 +8,6 @@ The [three outputs](device_output.npz) were finite FP32 hidden/new-key/new-value
 
 All 73 execution-detail rows in the profile were attributed to NPU. The 100 component samples had nearest-rank p50/p95 **1.837/2.079 ms** (minimum 1.749 ms); the service's `estimated_inference_time` was that minimum, not the median. Reported component inference peak memory was 90,877,952 bytes. These hosted component measurements exclude prefill, 27 other decoder layers, sampling, handoff, loading and a resident KV/ring/controller. They are not paired speed comparisons with other devices or OS builds.
 
-The [audit script](../../../../experiments/audit_spark_qualcomm_attention.py) reproduces the provenance and numerical checks using pinned source artifact paths in the S24 record. The [collector](../../../../experiments/collect_spark_aihub_component.py) retrieves these existing jobs via the local client configuration; neither script contains credentials.
+The [audit script](../../../../experiments/audit_spark_qualcomm_attention.py) reproduces the provenance and numerical checks using pinned source artifact paths in the S24 record. The [collector](../../../../experiments/collect_aihub_component.py) retrieves these existing jobs via the local client configuration; neither script contains credentials.
 
 **Matrix status: component NPU inference/profiled; NOT E2E.** Full qualification needs a complete device-resident Spark artifact and Omni backend with prefill/decode/KV-ring/sampling, state, token quality, cancellation and memory admission, followed by sustained power and thermal measurement.

benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md

@@ -8,6 +8,6 @@ The [three outputs](device_output.npz) were finite FP32 hidden/new-key/new-value
 
 All 73 execution-detail rows in the device profile were attributed to NPU. The 100 component inference-time samples had nearest-rank p50/p95 **1.433/1.606 ms** (minimum 1.346 ms). The service's `estimated_inference_time` was 1.346 ms, matching the minimum rather than the median. The reported component inference peak memory was 18,423,808 bytes. These are hosted component-job measurements; they exclude prefill, 27 other decoder layers, sampling, transfer, loading and a resident KV/ring/controller. Hardware, OS and hosted power state differ from the mobile runs, so these are not paired speed comparisons.
 
-Re-audit with `benchmarks/edge_harness/experiments/audit_spark_qualcomm_attention.py` and the [pinned source artifact paths](../spark_s24_full_attention/README.md). `benchmarks/edge_harness/experiments/collect_spark_aihub_component.py` retrieves the existing jobs using the local AI Hub client configuration. Neither script contains credentials.
+Re-audit with `benchmarks/edge_harness/experiments/audit_spark_qualcomm_attention.py` and the [pinned source artifact paths](../spark_s24_full_attention/README.md). `benchmarks/edge_harness/experiments/collect_aihub_component.py` retrieves the existing jobs using the local AI Hub client configuration. Neither script contains credentials.
 
 **Matrix status: component NPU inference/profiled; NOT E2E.** Full qualification still needs an X Elite machine running the complete 28-layer Spark model through a resident Omni backend with prefill/decode/KV-ring/sampling, state and cancellation, token-quality checks, memory admission and sustained power/thermal measurements. Workbench is a test facility, not that deployment runtime.

benchmarks/edge_harness/results/e2e_expansion_20260924/README.md

@@ -0,0 +1,3 @@
+# 2026-09-24 Qualcomm component expansion
+
+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 inference produces a saturated waveform, so its historical profile must not be interpreted as a usable TTS path. Workbench jobs for RB3, X Elite and SA8775P controls were still pending when this snapshot was written; their submission IDs are retained for completion. 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

@@ -0,0 +1,13 @@
+# Qwen3-TTS Code2Wav component across Qualcomm devices
+
+**The retained Qwen3-TTS Code2Wav export does not yet have a numerically qualified SA8775P GPU path.** A fresh inference on the exact hosted SA8775P ADP returned a severely saturated waveform on the same input for which original ONNX Runtime CPU and prior S25 GPU outputs remained unsaturated. Other submitted device jobs were pending at this evidence snapshot; they are not counted as passes.
+
+The historical Workbench source model `mqyer339n` is a 423,655,047-byte ONNX graph with one FP32 input `[1,512,97]` and one 48,000-sample output. The [manifest](artifact_manifest.json) pins its SHA256, the retained Workbench fixture dataset `d7m80jl32`, a local ONNX Runtime 1.29 CPU output, compiled target downloads, and the prior [S25 GPU output](s25_reference_output.npz). The source export script identifies the Qwen3-TTS 0.6B CustomVoice model family, but did not attest the exact checkpoint revision; these results must not be promoted to a revision-qualified release claim. Large source/target artifacts and the fixture remain outside Git under `/home/zhout/project/edge_infer/models/`.
+
+The [S25 historical inference](s25_reference_inference.json) on the same fixture differed from local FP32 ONNX CPU by **4.24% relative L2 / 27.46 dB SNR** ([comparison](s25_cpu_comparison.json)); that waveform has no listening-quality tolerance. A fresh [SA8775P inference](sa8775p/inference_report.json) used historical compiled TFLite model `mnze95don` and requested GPU. Its [output](sa8775p/device_output.npz) had **96.63%** of samples at absolute amplitude at least 0.999, RMS **0.989** versus CPU **0.0837**, relative L2 **11.832**, and SNR **−21.46 dB**. The fresh [profile](sa8775p/profile_report.json) attributed 403 execution-detail rows to CPU and 335 to GPU; 100 component samples had nearest-rank p50/p95 **1.776/1.880 s** (minimum 1.675 s). The [audit](sa8775p/audit_report.json) pins all output hashes, waveform differences and placement. This is measured execution of a **numerically failed** output, not a usable TTS performance result. The [SA8775P compile record](sa8775p/compile_report.json) pins the artifact and source IDs.
+
+The re-fetched *historical* [SA8775P profile](sa8775p/historical_profile_report.json) has 100 component samples, nearest-rank p50/p95 **1.751/1.874 s**, and 403 CPU plus 335 GPU execution-detail rows. The *historical* [RB3 profile](rb3/historical_profile_report.json) has 100 samples, p50/p95 **3.143/3.189 s**, also with 403 CPU plus 335 GPU rows. Their previously quoted 1.701/3.107 s estimated-time fields are sample minima, not medians. Both profiles exclude talker/predictor, stage handoff, streaming and full-request time. The SA and RB3 TFLite target downloads are bitwise identical, but that does not establish matching runtime outputs or quality.
+
+Current [RB3](rb3/inference_submission.json) fresh job submissions, [SA CPU](sa8775p/cpu_control/inference_submission.json) and [FP32-preserving GPU](sa8775p/gpu_fp32/inference_submission.json) controls, and X Elite [DirectML GPU](xelite/gpu/inference_submission.json) and [CPU](xelite/cpu/inference_submission.json) submissions remain pending in this snapshot. The [X Elite ONNX compile](xelite/compile_report.json) passed after the older TFLite target was rejected on Windows ARM. Pending jobs have no numerical or latency conclusion. The [collector](../../../../experiments/collect_aihub_component.py) retrieves their terminal reports without resubmitting jobs, and the [audit script](../../../../experiments/audit_qwen_tts_vocoder_qualcomm.py) checks source/fixture hashes, device and model IDs, actual placement, waveform error, saturation and sample percentiles.
+
+This is one fixed vocoder component with a retained input, not complete Qwen3-TTS. The next gates are a quality-passing SA/RB3/X Elite target, talker and code predictor numerical checks, device-resident handoff/co-residency, playable streaming, cancellation, memory admission and sustained power/thermal profiling.

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/artifact_manifest.json

@@ -0,0 +1,79 @@
+{
+  "facility": "Qualcomm AI Hub Workbench",
+  "source_model_ids": {
+    "talker": "mq3xre4rq",
+    "vocoder": "mqyer339n"
+  },
+  "fixture_dataset_id": "d7m80jl32",
+  "reference_inference_job_id": "jpx428ojg",
+  "checkpoint_revision": "not attested by historical export; component provenance only",
+  "onnxruntime_cpu_version": "1.29.0",
+  "files": {
+    "talker_source_onnx": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_talker_step2_L256_20260924.onnx",
+      "bytes": 1776377476,
+      "sha256": "41feb08517678e94d9301e992d0df610fc60e7c9faba5600703298c86554fd22"
+    },
+    "vocoder_source_onnx": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_code2wav_c25_20260924.onnx",
+      "bytes": 423655047,
+      "sha256": "7c07666229c6d404894132e22f784525e3212a49efad55ddd532eda5286c2b2b"
+    },
+    "vocoder_fixture": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_vocoder_fixture_20260924.npz",
+      "bytes": 129721,
+      "sha256": "dfca9e8d2a724a568a31b6cd18ff6a52fdc205fd34e13cbf1cd75f0f7937e82d"
+    },
+    "vocoder_ort_cpu_output": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_vocoder_ort_cpu_output_20260924.npz",
+      "bytes": 179942,
+      "sha256": "8de78b94b59f01e49852f7706adbd2508a9976b8e2705cfd3b94d07bb3606cb1"
+    },
+    "s25_historical_gpu_output": {
+      "path": "/home/zhout/project/edge_infer/vllm-omni-edge/benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25_reference_output.npz",
+      "bytes": 115235,
+      "sha256": "4a26486d5e84a6e22d16f9b26f995b8fe645c29bbd70b103615ec35a0fcb10a0"
+    },
+    "sa8775p_tflite": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_vocoder_sa8775p_20260924.tflite",
+      "bytes": 421658132,
+      "sha256": "29826da6ecc1ec2efc0e3589c3cebb42b20d7e20c0975a037d9f51625d6638cd",
+      "model_id": "mnze95don",
+      "compile_job_id": "jp410o3vp"
+    },
+    "rb3_tflite": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_vocoder_rb3_20260924.tflite",
+      "bytes": 421658132,
+      "sha256": "29826da6ecc1ec2efc0e3589c3cebb42b20d7e20c0975a037d9f51625d6638cd",
+      "model_id": "mn0wyzd3q",
+      "compile_job_id": "j5mmy28w5"
+    },
+    "xelite_onnx": {
+      "path": "/home/zhout/project/edge_infer/models/qwen3tts_vocoder_xelite_20260924.onnx.onnx.zip",
+      "bytes": 394632184,
+      "sha256": "e8f80a85d6f36248ad015edf63e553514011ae621d234a1e3d88ea71f93568f0",
+      "model_id": "mnz079ezm",
+      "compile_job_id": "j5qlyz8op"
+    }
+  },
+  "fixture": {
+    "shape": [
+      1,
+      512,
+      97
+    ],
+    "dtype": "float32",
+    "min": -90.89772033691406,
+    "max": 65.62191772460938,
+    "mean": 0.0705118179321289,
+    "std": 11.206748962402344
+  },
+  "cpu_reference": {
+    "shape": [
+      1,
+      48000
+    ],
+    "dtype": "float32",
+    "rms": 0.08373439311981201
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/rb3/compile_report.json

@@ -0,0 +1,31 @@
+{
+  "retrieved_utc": "2026-09-23T22:21:24.623865+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5mmy28w5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5mmy28w5/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mqyer339n",
+  "target_model_id": "mn0wyzd3q",
+  "device": {
+    "name": "Dragonwing RB3 Gen 2 Vision Kit",
+    "os": "1.6",
+    "attributes": [
+      "os:qc_linux",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-oe-linux-gcc11",
+      "format:iot",
+      "chipset:qualcomm-qcs6490",
+      "chipset:qcs6490",
+      "chipset:qualcomm-dragonwing-qcs6490",
+      "hexagon:v68",
+      "soc-model:93",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime tflite",
+  "hub_version": "aihub-2026.08.28.1"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/rb3/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:21:06.218268+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5mmy28w5",
+  "target_model_id": "mn0wyzd3q",
+  "input_dataset_id": "d7m80jl32",
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit gpu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jg9zmkv8p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jg9zmkv8p/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/rb3/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:21:07.854394+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5mmy28w5",
+  "target_model_id": "mn0wyzd3q",
+  "input_dataset_id": null,
+  "device": "Dragonwing RB3 Gen 2 Vision Kit",
+  "options": "--compute_unit gpu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jp1nj707g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp1nj707g/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25_cpu_comparison.json

@@ -0,0 +1,16 @@
+{
+  "scope": "historical S25 TFLite GPU inference on same retained vocoder fixture; component comparison only",
+  "source_model_id": "mqyer339n",
+  "source_cpu_onnxruntime_version": "1.29.0",
+  "s25_inference_job_id": "jpx428ojg",
+  "input_dataset_id": "d7m80jl32",
+  "status": "component_waveform_quality_unqualified",
+  "cpu_source_vs_s25_gpu": {
+    "relative_l2": 0.04238531411741596,
+    "snr_db": 27.455691878796923,
+    "max_abs": 0.10591871291399002,
+    "cosine": 0.9991021814913129,
+    "reference_rms": 0.08373439205485074,
+    "output_rms": 0.08376787228276501
+  }
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/s25_reference_inference.json

@@ -0,0 +1,32 @@
+{
+  "retrieved_utc": "2026-09-23T22:23:04.763805+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jpx428ojg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpx428ojg/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "model_id": "mng7kwxon",
+  "input_dataset_id": "d7m80jl32",
+  "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"
+    ]
+  },
+  "options": "--compute_unit gpu",
+  "hub_version": "aihub-2026.08.28.1"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/audit_report.json

@@ -0,0 +1,86 @@
+{
+  "scope": "one Qwen3-TTS code2wav 48,000-sample component on one retained fixture; no complete TTS stream",
+  "status": "component_executed_quality_unqualified",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "compile_job_id": "jp410o3vp",
+  "inference_job_id": "jgzlm8ex5",
+  "profile_job_id": "j5wl78omp",
+  "target_model_id": "mnze95don",
+  "input_dataset_id": "d7m80jl32",
+  "s25_prior_inference_job_id": "jpx428ojg",
+  "requested_run_options": "--compute_unit gpu",
+  "files": {
+    "source_onnx": {
+      "bytes": 423655047,
+      "sha256": "7c07666229c6d404894132e22f784525e3212a49efad55ddd532eda5286c2b2b"
+    },
+    "fixture": {
+      "bytes": 129721,
+      "sha256": "dfca9e8d2a724a568a31b6cd18ff6a52fdc205fd34e13cbf1cd75f0f7937e82d"
+    },
+    "cpu_output": {
+      "bytes": 179942,
+      "sha256": "8de78b94b59f01e49852f7706adbd2508a9976b8e2705cfd3b94d07bb3606cb1"
+    },
+    "s25_output": {
+      "bytes": 115235,
+      "sha256": "4a26486d5e84a6e22d16f9b26f995b8fe645c29bbd70b103615ec35a0fcb10a0"
+    },
+    "device_output": {
+      "bytes": 10189,
+      "sha256": "770c9c1f5a8dcaf9d8054cd2617bb8cd47d0c005c6858836452882ff3d7353d5"
+    }
+  },
+  "cpu_source_vs_device": {
+    "relative_l2": 11.832417281859046,
+    "snr_db": -21.461469541638177,
+    "max_abs": 1.4882186353206635,
+    "cosine": 0.017841868329608154,
+    "reference_rms": 0.08373439205485074,
+    "output_rms": 0.988730688363501,
+    "output_saturated_fraction_abs_ge_0_999": 0.9663333333333334
+  },
+  "s25_gpu_vs_device": {
+    "relative_l2": 11.82796496222066,
+    "snr_db": -21.458200587056638,
+    "max_abs": 1.488037109375,
+    "cosine": 0.01759117063047259,
+    "reference_rms": 0.08376787228276501,
+    "output_rms": 0.988730688363501,
+    "output_saturated_fraction_abs_ge_0_999": 0.9663333333333334
+  },
+  "profile": {
+    "sample_count": 100,
+    "sample_min_us": 1674704,
+    "nearest_rank_p50_us": 1775673,
+    "nearest_rank_p95_us": 1880119,
+    "reported_estimated_inference_time_us": 1674704,
+    "reported_peak_memory_bytes": 361168896,
+    "compute_unit_row_counts": {
+      "CPU": 403,
+      "GPU": 335
+    }
+  },
+  "limits": [
+    "Historical source export does not attest its exact checkpoint revision.",
+    "This fixed component waveform has no listening or task-quality tolerance.",
+    "Hosted component profiling excludes talker, predictor, stage handoff, streaming, admission and sustained power/thermal behavior."
+  ]
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/compile_report.json

@@ -0,0 +1,31 @@
+{
+  "retrieved_utc": "2026-09-23T22:21:23.511773+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jp410o3vp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp410o3vp/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mqyer339n",
+  "target_model_id": "mnze95don",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime tflite",
+  "hub_version": "aihub-2026.08.28.1"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/cpu_control/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:34:29.858166+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jp410o3vp",
+  "target_model_id": "mnze95don",
+  "input_dataset_id": "d7m80jl32",
+  "device": "SA8775P ADP",
+  "options": "--compute_unit cpu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jp1nj7x7g",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jp1nj7x7g/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/gpu_fp32/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:34:32.086663+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jp410o3vp",
+  "target_model_id": "mnze95don",
+  "input_dataset_id": "d7m80jl32",
+  "device": "SA8775P ADP",
+  "options": "--compute_unit gpu --tflite_options allow_fp32_as_fp16=false",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jgdd38lzg",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgdd38lzg/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/inference_report.json

@@ -0,0 +1,32 @@
+{
+  "timestamp_utc": "2026-09-23T22:33:06.275073+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "jgzlm8ex5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgzlm8ex5/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "options": "--compute_unit gpu",
+  "hub_version": "aihub-2026.09.11.0",
+  "device": {
+    "name": "SA8775P ADP",
+    "os": "14",
+    "attributes": [
+      "os:android",
+      "framework:tflite",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-android",
+      "format:auto",
+      "chipset:qualcomm-sa8775p",
+      "hexagon:v73",
+      "soc-model:52",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "model_id": "mnze95don",
+  "input_dataset_id": "d7m80jl32",
+  "outputs": "device_output.npz"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:21:01.692875+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jp410o3vp",
+  "target_model_id": "mnze95don",
+  "input_dataset_id": "d7m80jl32",
+  "device": "SA8775P ADP",
+  "options": "--compute_unit gpu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jgzlm8ex5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgzlm8ex5/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/sa8775p/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:21:03.431758+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "jp410o3vp",
+  "target_model_id": "mnze95don",
+  "input_dataset_id": null,
+  "device": "SA8775P ADP",
+  "options": "--compute_unit gpu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "j5wl78omp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl78omp/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/compile_report.json

@@ -0,0 +1,31 @@
+{
+  "retrieved_utc": "2026-09-23T22:27:16.875471+00:00",
+  "qai_hub_version": "0.55.0",
+  "facility": "Qualcomm AI Hub Workbench",
+  "job_id": "j5qlyz8op",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5qlyz8op/",
+  "status": "SUCCESS",
+  "status_message": "",
+  "source_model_id": "mqyer339n",
+  "target_model_id": "mnz079ezm",
+  "device": {
+    "name": "Snapdragon X Elite CRD",
+    "os": "11",
+    "attributes": [
+      "os:windows",
+      "format:compute",
+      "framework:onnx",
+      "vendor:qualcomm",
+      "abi:aarch64-windows",
+      "chipset:qualcomm-snapdragon-x-elite",
+      "chipset:sc8380xp",
+      "hexagon:v73",
+      "soc-model:60",
+      "htp-supports-fp16:true",
+      "htp-supports-weight-sharing:true",
+      "framework:qnn"
+    ]
+  },
+  "options": "--target_runtime onnx",
+  "hub_version": "aihub-2026.09.11.0"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/compile_submission.json

@@ -0,0 +1,11 @@
+{
+  "timestamp_utc": "2026-09-23T22:24:46.161553+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "job_id": "j5qlyz8op",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5qlyz8op/",
+  "source_model_id": "mqyer339n",
+  "device": "Snapdragon X Elite CRD",
+  "options": "--target_runtime onnx",
+  "reason": "historical TFLite target rejected on Windows ARM; try ONNX runtime path"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/cpu/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:27:38.212208+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5qlyz8op",
+  "target_model_id": "mnz079ezm",
+  "input_dataset_id": "d7m80jl32",
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit cpu",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "j5wl78nmp",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/j5wl78nmp/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/cpu/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:27:39.866884+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5qlyz8op",
+  "target_model_id": "mnz079ezm",
+  "input_dataset_id": null,
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit cpu",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jg9zmke8p",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jg9zmke8p/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/gpu/inference_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:27:34.392680+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5qlyz8op",
+  "target_model_id": "mnz079ezm",
+  "input_dataset_id": "d7m80jl32",
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit gpu --onnx_execution_providers directml",
+  "kind": "inference",
+  "status": "SUBMITTED",
+  "job_id": "jpe7z83v5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jpe7z83v5/"
+}

benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/xelite/gpu/profile_submission.json

@@ -0,0 +1,14 @@
+{
+  "timestamp_utc": "2026-09-23T22:27:36.042380+00:00",
+  "facility": "Qualcomm AI Hub Workbench",
+  "qai_hub_version": "0.55.0",
+  "source_compile_job_id": "j5qlyz8op",
+  "target_model_id": "mnz079ezm",
+  "input_dataset_id": null,
+  "device": "Snapdragon X Elite CRD",
+  "options": "--compute_unit gpu --onnx_execution_providers directml",
+  "kind": "profile",
+  "status": "SUBMITTED",
+  "job_id": "jgzlm8kx5",
+  "url": "https://workbench.aihub.qualcomm.com/jobs/jgzlm8kx5/"
+}

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

@@ -11,6 +11,8 @@ The [WSL CPU Spark GGUF Omni profile](../../../e2e_expansion_20260923/evidence/s
 On a hosted [Galaxy S24 Spark attention-layer experiment](../../../e2e_expansion_20260923/evidence/spark_s24_full_attention/README.md), a real one-layer W8A16 QDQ artifact executed on NPU and matched prior S25 component outputs exactly on one dataset. The 100-sample layer p50/p95 was 1.109/1.205 ms; the service's 1.098 ms estimated-time field was the sample minimum. FP32-source versus S24 output hidden-state relative L2 was 1.351% without a downstream token-quality gate. This adds component C evidence only; device-local generation remains absent.
 
 The same W8A16 layer also ran on hosted [Snapdragon X Elite CRD](../../../e2e_expansion_20260923/evidence/spark_xelite_full_attention/README.md) and [SA8775P ADP](../../../e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md) NPUs, with pinned-fixture outputs identical to S24/S25; the 100-sample component p50/p95 was 1.433/1.606 ms and 1.837/2.079 ms, respectively. On [RB3 Gen 2 / QCS6490](../../../e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md), that W8A16 graph failed QNN DLC loading even after target-specific recompilation. An FP32 ONNX CPU layer then passed with source-relative L2 below 9e-7, while a calibrated W8A8 NPU layer executed but had 20.9% hidden-state relative L2 and remains numerically unqualified. These hosted component outcomes do not establish full generation or comparable device speed.
+
+A [Qwen3-TTS Code2Wav follow-up](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md) recovered source and target artifacts, a retained fixture, and raw historical profiles. On SA8775P, fresh GPU-requested inference saturated 96.63% of waveform samples and differed from original ONNX CPU by 11.832 relative L2; this tested vocoder path is numerically unqualified. Historical SA8775P/RB3 profile p50/p95 values were 1.751/1.874 s and 3.143/3.189 s, respectively, with both CPU and GPU rows; earlier estimated-time fields were minima. X Elite ONNX compile passed after the old TFLite rejection. Fresh profile/control/other-device jobs remain pending, so they have no new execution conclusion.
 
 ## Configuration × model
 
@@ -77,7 +79,7 @@ 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 | NOT_E2E_PROFILEABLE | Historical QNN talker profile re-fetched successfully; no numerical/full-pipeline gate. Historical TFLite vocoder target was refused on this Windows ARM device. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
+| PC Snapdragon X Elite CRD / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_ONNX_VOCODER_COMPILE_PASS; NOT_E2E_PROFILEABLE | Historical QNN talker profiling exists without a numerical or full-pipeline gate. The older TFLite vocoder target was rejected on this Windows ARM device; a fresh ONNX vocoder compile passed, but its explicit DirectML GPU and CPU inference/profile jobs were pending at this snapshot. No waveform or actual placement conclusion follows from compile alone. | Retrieve and audit the pending ONNX vocoder jobs, qualify talker/predictor and complete local streaming, memory and sustained behavior; [vocoder evidence](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md). |
 | 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 | 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 | InternVLA-A1 | 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. |
@@ -92,12 +94,12 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | Historical FP16 speech-head job re-fetched successfully; SNR 49.22 dB/top-1 match in the recorded parity test. W8A16 top-1 failed. Not complete MiniCPM-o. | Qualify thinker/encoders/vocoder and stateful local integration. |
 | Mobile Galaxy S24 / Snapdragon 8 Gen 3 | InternVLA-A1 | 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. |
 | Embedded SA8775P ADP / 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 SA8775P ADP (Android 14). 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 1.837/2.079 ms; the 1.749 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 embedded resident artifact/controller with state, token quality, memory admission/cancel and sustained power; [SA8775P NPU raw evidence](../../../e2e_expansion_20260923/evidence/spark_sa8775_full_attention/README.md). |
-| Embedded SA8775P ADP / AI Hub | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical NPU talker and GPU vocoder profiles re-fetched successfully; no complete pipeline, parity or co-residency qualification. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
+| Embedded SA8775P ADP / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_VOCODER_NUMERIC_FAIL; NOT_E2E_PROFILEABLE | Historical NPU talker and mixed CPU+GPU vocoder profiles exist without a complete pipeline. A fresh GPU-requested Code2Wav inference on the retained fixture saturated 96.63% of waveform samples, with 11.832 relative L2 and -21.46 dB SNR versus original ONNX CPU; this tested vocoder path is numerically unqualified. The fresh profile has 100 samples at nearest-rank p50/p95 1.776/1.880 s with 403 CPU and 335 GPU execution-detail rows, not an all-GPU stage; it times a numerically failed output. CPU/FP32-GPU control jobs were pending. | Isolate delegate/precision failure, establish waveform and listening quality, then validate co-resident talker/predictor/vocoder, streaming, memory and sustained behavior; [raw vocoder evidence](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md). |
 | Embedded SA8775P ADP / 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. |
 | Embedded SA8775P ADP / AI Hub | MiniCPM-o 4.5 | 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. |
 | Embedded SA8775P ADP / AI Hub | InternVLA-A1 | 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. |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | Spark-X2.5 | COMPONENT_CPU_PARITY_PASS; COMPONENT_NPU_NUMERIC_UNQUALIFIED; NOT_E2E_PROFILEABLE | On the exact hosted Dragonwing RB3 Gen 2 Vision Kit (QCS6490, QC Linux 1.6), the reused S24 W8A16 QNN DLC and a separately RB3-compiled W8A16 DLC each failed inference/profile at QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR. An FP32 ONNX one-layer CPU route passed with all 56 profile rows on CPU, source-relative hidden/K/V L2 of 8.73e-7/2.98e-7/3.37e-7 and 100-sample component p50/p95 19.304/21.113 ms. A calibrated W8A8 QNN DLC ran with all 61 rows on NPU and p50/p95 5.610/5.761 ms, but hidden/K/V relative L2 was 20.914%/1.702%/4.413% versus FP32 CPU, without a next-token quality gate. These are different precisions, not a qualified speedup; no complete 28-layer prefill/decode/KV-ring/sampling, local Omni controller or complete-request timing exists. | Improve representative W8A8 calibration or find another compatible quality-passing artifact; validate token parity, then full device-resident generation, admission/cancel and sustained power; [RB3 raw evidence](../../../e2e_expansion_20260923/evidence/spark_rb3_full_attention/README.md). |
-| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | Historical GPU vocoder profile re-fetched successfully. The attempted NPU talker build was rejected for floating-point inputs. No full model pipeline. | Validate target-specific numerical quality, complete local execution, memory and sustained behavior. |
+| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_VOCODER_PROFILE_ONLY; NOT_E2E_PROFILEABLE | The historical TFLite vocoder profile re-fetched with 100 samples at nearest-rank p50/p95 3.143/3.189 s and 403 CPU plus 335 GPU execution-detail rows. Its estimated-time field 3.107 s is the minimum, not median. Fresh same-fixture inference/profile was pending, so numerical quality is unknown. The historical NPU talker build was rejected for floating-point inputs; no full model pipeline exists. | Retrieve fresh vocoder output and placement, establish a quality-passing integer talker artifact, then validate complete local execution, memory and sustained behavior; [vocoder evidence](../../../e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md). |
 | Embedded RB3 Gen 2 / QCS6490 / 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. |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | MiniCPM-o 4.5 | 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. |
 | Embedded RB3 Gen 2 / QCS6490 / AI Hub | InternVLA-A1 | 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. |

docs/design/omni_edge_e2e_roadmap.md

@@ -15,6 +15,8 @@ remain open; the 2026-09-22 audit remains the historical baseline.
 
 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.
 
+The [Qwen3-TTS Code2Wav Qualcomm follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260924/evidence/qwen_tts_vocoder_qualcomm/README.md) adds a retained same-fixture numerical check to M2. A fresh SA8775P GPU-requested output saturated 96.63% of samples and differs from original ONNX CPU by 11.832 relative L2, so the tested vocoder target fails the numerical gate despite historical component profiling. Historical SA8775P and RB3 vocoder profiles have mixed CPU/GPU placement; their estimated-time fields were minima. X Elite ONNX compile passed, while RB3/X Elite runtime and SA control jobs remained pending at this snapshot. M2 still requires a quality-passing talker/predictor/vocoder set, device-local co-residency and handoff, playable streaming and sustained memory/power validation.
+
 The [WSL CPU expansion](../../benchmarks/edge_harness/results/e2e_expansion_20260923/README.md)
 adds constrained three-stage MiniCPM-o text-to-speech requests and a CPU-only
 InternVLA synthetic policy forward. A subsequent 20-request serial MiniCPM-o