Qualify scoped MiniCPM-o Radeon hybrid path

This commit is contained in:
tzhouam committed 2026-09-23 22:43:20 +08:00
1 parent 445ca92ef1
commit 13cc723c71
150 files changed
+6054 -9

No files matched your search

@@ -0,0 +1,86 @@
#!/usr/bin/env python3
"""Verify the same prompted MiniCPM-o request across CPU and Radeon runs."""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
def read_json(path: Path) -> dict:
return json.loads(path.read_text(encoding="utf-8"))
def language_tokens(path: Path) -> list[int]:
matches = re.findall(r"LLM->TTS:.*?token_ids=\[([^]]+)\]", path.read_text(
encoding="utf-8", errors="replace"
))
if len(matches) != 1:
raise RuntimeError(f"expected exactly one logged text chunk: {path}")
return [int(value) for value in matches[0].split()]
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--cpu-dir", type=Path, required=True)
parser.add_argument("--radeon-dir", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
cpu, gpu = (read_json(path / "full_report.json") for path in
(args.cpu_dir, args.radeon_dir))
if (cpu["expected_placement"], gpu["expected_placement"]) != (
"cpu", "radeon-hybrid"
):
raise RuntimeError("incorrect device placement")
if any(item["status"] != "scoped_complete_path_pass" for item in (cpu, gpu)):
raise RuntimeError("one complete path failed")
for key in ("artifacts", "model_revision", "input_image_sha256",
"input_audio_sha256", "response_text"):
if cpu[key] != gpu[key]:
raise RuntimeError(f"cross-device mismatch: {key}")
cpu_tokens = language_tokens(args.cpu_dir / "full.stdout.log")
gpu_tokens = language_tokens(args.radeon_dir / "full.stdout.log")
if cpu_tokens != gpu_tokens:
raise RuntimeError("language token IDs differ")
input_asr = read_json(args.radeon_dir / "input_asr.json")
cpu_asr = read_json(args.cpu_dir / "output_asr.json")
gpu_asr = read_json(args.radeon_dir / "output_asr.json")
if input_asr["word_error_rate"] != 0 or any(
item["word_error_rate"] != 0 for item in (cpu_asr, gpu_asr)
):
raise RuntimeError("input or output speech failed the ASR proxy")
if cpu["response_text"] != "The square in the image is red.":
raise RuntimeError("visual question was not answered correctly")
report = {
"scope": "one synthetic spoken visual question; standalone CPU vs Radeon hybrid",
"status": "scoped_prompted_path_pass",
"checkpoint_revision": cpu["model_revision"],
"same_artifact_set": True,
"same_input_image_audio": True,
"spoken_input": input_asr["reference_text"],
"input_asr_wer": input_asr["word_error_rate"],
"response_text": cpu["response_text"],
"language_token_ids": cpu_tokens,
"cpu": {
"cold_wall_s": cpu["sampled_process_profile"]["wall_s"],
"audio_duration_s": cpu["audio_duration_s"],
"pcm_sha256": cpu["pcm_sha256"],
"output_asr_wer": cpu_asr["word_error_rate"],
},
"radeon_hybrid": {
"cold_wall_s": gpu["sampled_process_profile"]["wall_s"],
"audio_duration_s": gpu["audio_duration_s"],
"pcm_sha256": gpu["pcm_sha256"],
"output_asr_wer": gpu_asr["word_error_rate"],
},
"quality_limit": "one synthetic red square and one spoken question",
}
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
@@ -52,6 +52,9 @@ def main() -> None:
parser.add_argument("--merged-wav", type=Path, required=True)
parser.add_argument("--report", type=Path, required=True)
parser.add_argument("--platform-label", default="WSL CPU")
parser.add_argument(
"--expected-placement", choices=("cpu", "radeon-hybrid"), default="cpu"
)
parser.add_argument("--process-profile", type=Path)
args = parser.parse_args()
@@ -59,15 +62,27 @@ def main() -> None:
log = args.run_log.read_text(encoding="utf-8", errors="replace")
required_log = (
"Media type: 2 (omni: audio+vision)",
"GPU layers: 0",
"Token2Wav device: cpu",
"vision_ctx: vision using CPU backend",
"Token2Wav: initialized successfully",
"stream_prefill(index=1): processing user audio:",
"stream_prefill(index=1): user audio embedding:",
"stream_prefill: prefilled 1 vision chunks",
)
missing_log = [item for item in required_log if item not in log]
placement_log = {
"cpu": (
"GPU layers: 0",
"vision_ctx: vision using CPU backend",
),
"radeon-hybrid": (
"GPU layers: 99",
"using device Vulkan0 (AMD Radeon(TM) 890M Graphics)",
"offloaded 37/37 layers to GPU",
"vision_ctx: vision using Vulkan0 backend",
"TTS model: loading with n_gpu_layers=0",
"Token2Wav: non-CUDA backend, vocoder using CPU",
),
}[args.expected_placement]
missing_log = [item for item in required_log + placement_log if item not in log]
if missing_log:
raise RuntimeError(f"missing complete-path/placement evidence: {missing_log}")
process_profile = None
@@ -138,6 +153,7 @@ def main() -> None:
report = {
"scope": f"standalone MiniCPM-o 4.5 GGUF audio+image input to text+speech on {args.platform_label}",
"status": "scoped_complete_path_pass",
"expected_placement": args.expected_placement,
"quality": "synthetic red-square/tone response not established as semantically correct",
"model_revision": args.model_revision,
"artifacts": artifacts,