Upload mi300x_9b/niah.py with huggingface_hub

mi300x_9b/niah.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python
+"""Minimal needle-in-a-haystack long-context retrieval check (stdlib only).
+
+Builds a long filler haystack, inserts a unique passcode needle at a chosen
+depth, asks for the passcode, and checks the greedy completion contains it.
+Used to gate the LCM page-unify split (L19 kept NATURAL (3 retr,1 windowed-local))
+against the phantom baseline (L19 DEMOTED to (4,0) all-retrieval) at the long
+contexts where the windowed-local heads actually matter.
+
+Usage (run on the GPU host, server on :8000):
+  python bench/niah.py --model rtp9 --tag lcm   --out /work/scratch/niah_lcm.json
+  python bench/niah.py --model rtp9 --tag phantom --out /work/scratch/niah_phantom.json
+  python bench/niah.py --summary /work/scratch/niah_lcm.json
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+import urllib.request
+
+# Filler sentence ~ a handful of tokens; repeated to reach target length.
+FILLER = (
+    "The city archives record the slow turning of the seasons over the river "
+    "delta, where merchants once traded salt and amber along the old stone road. "
+)
+NEEDLE = "The secret passcode for the vault is {code}. Remember it well. "
+QUESTION = (
+    "\n\nQuestion: What is the secret passcode for the vault? "
+    "Answer with only the number.\nAnswer:"
+)
+
+
+def gen(url, model, prompt, n_decode):
+    body = json.dumps({
+        "model": model,
+        "prompt": prompt,
+        "max_tokens": n_decode,
+        "temperature": 1.0,
+        "top_p": 0.95,
+        "top_k": 20,
+        "min_p": 0.0,
+        "presence_penalty": 1.5,
+        "repetition_penalty": 1.0,
+        "seed": 0,
+        "ignore_eos": False,
+        "stream": False,
+    }).encode("utf-8")
+    req = urllib.request.Request(
+        url, data=body,
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=1800) as resp:
+        return json.loads(resp.read().decode("utf-8"))
+
+
+def build_prompt(approx_tokens, depth_frac, code):
+    # ~4 chars/token heuristic; FILLER is ~30 tokens.
+    target_chars = approx_tokens * 4
+    needle = NEEDLE.format(code=code)
+    body_chars = max(0, target_chars - len(needle) - len(QUESTION))
+    n_filler = max(1, body_chars // len(FILLER))
+    pre_n = int(n_filler * depth_frac)
+    post_n = n_filler - pre_n
+    pre = FILLER * pre_n
+    post = FILLER * post_n
+    return f"{pre}{needle}{post}{QUESTION}"
+
+
+def run(args):
+    code = "738291"
+    lengths = [int(x) for x in args.lengths.split(",")]
+    depths = [float(x) for x in args.depths.split(",")]
+    results = []
+    for L in lengths:
+        for d in depths:
+            prompt = build_prompt(L, d, code)
+            resp = gen(args.url, args.model, prompt, args.decode_tokens)
+            choice = resp["choices"][0]
+            text = choice.get("text", "")
+            ptok = resp.get("usage", {}).get("prompt_tokens")
+            ok = code in text
+            results.append({
+                "approx_len": L, "prompt_tokens": ptok,
+                "depth": d, "found": ok, "out": text[:80],
+            })
+            print(f"[{args.tag}] len~{L} ptok={ptok} depth={d:.2f} "
+                  f"found={ok} out={text[:48]!r}")
+    rec = {"tag": args.tag, "code": code, "results": results}
+    if args.out:
+        with open(args.out, "w", encoding="utf-8") as f:
+            json.dump(rec, f)
+    n_ok = sum(r["found"] for r in results)
+    print(f"[{args.tag}] PASS {n_ok}/{len(results)}")
+    return 0
+
+
+def summary(path):
+    rec = json.load(open(path, encoding="utf-8"))
+    rs = rec["results"]
+    n_ok = sum(r["found"] for r in rs)
+    print(f"[{rec['tag']}] PASS {n_ok}/{len(rs)}")
+    for r in rs:
+        mark = "OK " if r["found"] else "MISS"
+        print(f"  {mark} ptok={r['prompt_tokens']} depth={r['depth']:.2f}")
+    return 0
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--url", default="http://localhost:8000/v1/completions")
+    ap.add_argument("--model", default="rtp9")
+    ap.add_argument("--lengths", default="8000,32000,64000,120000")
+    ap.add_argument("--depths", default="0.1,0.5,0.9")
+    ap.add_argument("--decode-tokens", type=int, default=8192)
+    ap.add_argument("--tag", default="run")
+    ap.add_argument("--out", default="")
+    ap.add_argument("--summary", default="")
+    args = ap.parse_args()
+    if args.summary:
+        return summary(args.summary)
+    return run(args)
+
+
+if __name__ == "__main__":
+    sys.exit(main())