Kev-4B-LiteRT: row-prefill graphs L512/L1024/L2048 (fp16 FC + int8 embedding), pointer head, tokenizer, Python host, fixtures, conversion scripts, card
9f1eee5 verified | """GPU-vs-CPU comparison of one file after both parity runs exist (used when the GPU run finished before the CPU run | |
| of the same file, so its parity json says gpu_vs_cpu_same_file.available = false). | |
| python gpu_vs_cpu.py --L 1024 --variant v2_fp16fc_i8emb --tag gpu_f32 [--model 4b] [--cache <dir>] [--cpu-sfx _smoke50] | |
| Reads results/litert_{tag}_rows_L{L}[_4b]_{variant}.json + <cache>/hsel_{tag}_L{L}[_4b]_{variant}.npz (GPU) and the CPU | |
| run's rows json + npz of the same file; writes results/litert_{tag}_vs_cpu_L{L}[_4b]_{variant}[{cpu-sfx}].json (never | |
| overwritten). --cpu-sfx _smoke50 compares with a CPU run of the first N questions (litert_parity.py --limit N); only | |
| the questions both runs have are compared.""" | |
| import argparse | |
| import json | |
| import numpy as np | |
| from r2_common import MODEL, RESULT_SUFFIX, K, add_model_arg, cache_dir, dump_json | |
| from r3_common import gpu_vs_cpu | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--L", type=int, required=True) | |
| ap.add_argument("--variant", required=True) | |
| ap.add_argument("--tag", choices=["gpu_f32", "gpu_f16"], required=True) | |
| ap.add_argument("--cache", default="", help="dir of the h_sel npz (default cache/r3 for 0.8B, cache/r4b for 4B)") | |
| ap.add_argument("--cpu-sfx", default="", help="suffix of a partial CPU run (e.g. _smoke50)") | |
| add_model_arg(ap) | |
| a = ap.parse_args() | |
| assert a.model == MODEL | |
| cache = (K / a.cache) if a.cache else cache_dir("cache/r3") | |
| sfx = f"{RESULT_SUFFIX}_{a.variant}" | |
| out = K / f"results/litert_{a.tag}_vs_cpu_L{a.L}{sfx}{a.cpu_sfx}.json" | |
| assert not out.exists(), f"refusing to overwrite {out.name}" | |
| rows = json.loads((K / f"results/litert_{a.tag}_rows_L{a.L}{sfx}.json").read_text())["rows"] | |
| g = np.load(cache / f"hsel_{a.tag}_L{a.L}{sfx}.npz") | |
| cpu_rows_path = K / f"results/litert_cpu_rows_L{a.L}{sfx}{a.cpu_sfx}.json" | |
| if a.cpu_sfx: # partial CPU run: compare only the questions it has | |
| have = {r["key"] for r in json.loads(cpu_rows_path.read_text())["rows"]} | |
| rows = [r for r in rows if r["key"] in have] | |
| doc = gpu_vs_cpu(rows, cpu_rows_path, {k: g[k] for k in g.files}, cache / f"hsel_cpu_L{a.L}{sfx}{a.cpu_sfx}.npz") | |
| doc = {"what": f"{a.tag} vs cpu, same file (L={a.L} {a.variant}, model {MODEL})", | |
| "gpu_parity_file": f"results/litert_{a.tag}_parity_L{a.L}{sfx}.json", **doc} | |
| dump_json(out, doc) | |
| print(json.dumps({k: v for k, v in doc.items() if k != "top10_by_dp"}, indent=1)) | |
| if __name__ == "__main__": | |
| main() | |