Upload folder using huggingface_hub

src/mathcompose/eval/run.py

@@ -33,12 +33,21 @@ def _runner(base_id, adapter):
     return HFRunner(base_id, adapter_path=adapter)
 
 
+def _prompts_per_batch(a):
+    from .runners import auto_concurrent_seqs
+    seqs = a.max_seqs or auto_concurrent_seqs()
+    prompts = max(1, seqs // max(1, a.maj_k))  # concurrent seqs = prompts * maj_k
+    print(f"[eval] maj_k={a.maj_k} concurrent_seqs~{seqs} prompts/batch={prompts} "
+          f"max_new_tokens={a.max_new_tokens}")
+    return prompts
+
+
 def cmd_processbench(a):
     from .processbench import evaluate_batched
     r = _runner(a.base_id, a.adapter)
-    bs = max(1, a.batch_size // max(1, a.maj_k))  # cap concurrent seqs = prompts * k
     res = evaluate_batched(r, splits=tuple(a.splits), limit=a.limit, n=a.maj_k,
-                           batch_size=bs, temperature=(0.6 if a.maj_k > 1 else 0.0),
+                           batch_size=_prompts_per_batch(a),
+                           temperature=(0.6 if a.maj_k > 1 else 0.0),
                            max_new_tokens=a.max_new_tokens)
     _save("processbench", a.tag, res)
 
@@ -46,8 +55,7 @@ def cmd_processbench(a):
 def cmd_prmbench(a):
     from .prmbench import evaluate_batched
     r = _runner(a.base_id, a.adapter)
-    bs = max(1, a.batch_size // max(1, a.maj_k))
-    res = evaluate_batched(r, limit=a.limit, n=a.maj_k, batch_size=bs,
+    res = evaluate_batched(r, limit=a.limit, n=a.maj_k, batch_size=_prompts_per_batch(a),
                            temperature=(0.6 if a.maj_k > 1 else 0.0),
                            max_new_tokens=a.max_new_tokens)
     _save("prmbench", a.tag, res)
@@ -107,7 +115,7 @@ def main() -> None:
     p.add_argument("--tag", default="tuned")
     p.add_argument("--maj-k", type=int, default=1)
     p.add_argument("--limit", type=int, default=None)
-    p.add_argument("--batch-size", type=int, default=16, help="max concurrent sequences (÷ maj-k)")
+    p.add_argument("--max-seqs", type=int, default=0, help="max concurrent sequences (0=auto by VRAM)")
     p.add_argument("--max-new-tokens", type=int, default=640)
     p.add_argument("--splits", nargs="+", default=["gsm8k", "math", "olympiadbench", "omnimath"])
     p.set_defaults(func=cmd_processbench)
@@ -117,7 +125,7 @@ def main() -> None:
     p.add_argument("--tag", default="tuned")
     p.add_argument("--maj-k", type=int, default=1)
     p.add_argument("--limit", type=int, default=None)
-    p.add_argument("--batch-size", type=int, default=16, help="max concurrent sequences (÷ maj-k)")
+    p.add_argument("--max-seqs", type=int, default=0, help="max concurrent sequences (0=auto by VRAM)")
     p.add_argument("--max-new-tokens", type=int, default=640)
     p.set_defaults(func=cmd_prmbench)
 

src/mathcompose/eval/runners.py

@@ -159,6 +159,23 @@ def make_solve_fn(runner: HFRunner, n: int = 1, temperature: float = 0.0,
     return solve_fn
 
 
+def auto_concurrent_seqs() -> int:
+    """Max sequences to generate at once, scaled to VRAM (eval is bf16, ~3GB weights
+    + KV cache, so a 1.5B model can run many in parallel)."""
+    import torch
+
+    if not torch.cuda.is_available():
+        return 8
+    vram = torch.cuda.get_device_properties(0).total_memory / 1e9
+    if vram >= 70:
+        return 128
+    if vram >= 40:
+        return 64
+    if vram >= 20:
+        return 32
+    return 12
+
+
 def verify_batch_predictions(runner: HFRunner, items, n: int = 1, batch_size: int = 16,
                              temperature: float = 0.0, max_new_tokens: int = 640,
                              desc: str = "verify") -> list[int]: