Path B v3: full test set + same-protocol refs (6.5482 / 7.3814)

quantization/hsaq/awq_path_b_clip.py

@@ -239,8 +239,10 @@ def apply_quantization(model, name_to_bits: dict[str, int],
 
 # ── Eval ────────────────────────────────────────────────────────────────
 
-def evaluate_ppl(model, tokenizer, ctx_len: int = 2048, n_samples: int = 40) -> float:
-    """Same wikitext PPL protocol as HSAQ baseline."""
+def evaluate_ppl(model, tokenizer, ctx_len: int = 2048, n_samples: int | None = None) -> tuple[float, int]:
+    """Same wikitext PPL protocol as awq_validation_runs.py.
+    n_samples=None evaluates FULL test set (per web-Claude review 2026-05-20).
+    Returns (ppl, n_windows_evaluated)."""
     from datasets import load_dataset
     ds = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
     text = "\n\n".join(ds["text"])
@@ -250,7 +252,9 @@ def evaluate_ppl(model, tokenizer, ctx_len: int = 2048, n_samples: int = 40) ->
     nlls = []
     stride = ctx_len
     prev_end = 0
-    for i in range(0, min(input_ids.size(1), n_samples * stride), stride):
+    max_pos = input_ids.size(1) if n_samples is None else min(input_ids.size(1), n_samples * stride)
+    n_windows = 0
+    for i in range(0, max_pos, stride):
         begin = max(i + stride - ctx_len, 0)
         end = min(i + stride, input_ids.size(1))
         trg_len = end - prev_end
@@ -261,7 +265,8 @@ def evaluate_ppl(model, tokenizer, ctx_len: int = 2048, n_samples: int = 40) ->
             out = model(ids, labels=target)
         nlls.append(out.loss.float() * trg_len)
         prev_end = end
-    return torch.exp(torch.stack(nlls).sum() / end).item()
+        n_windows += 1
+    return torch.exp(torch.stack(nlls).sum() / end).item(), n_windows
 
 # ── Upload ──────────────────────────────────────────────────────────────
 
@@ -358,13 +363,15 @@ def main():
         logger.info("Stage 5/5: PPL eval")
         if tokenizer.pad_token is None:
             tokenizer.pad_token = tokenizer.eos_token
-        ppl = evaluate_ppl(model, tokenizer)
+        ppl, n_windows = evaluate_ppl(model, tokenizer)
+        # bf16 / HSAQ refs from awq-validation-20260520_122419 (same protocol, full test set)
         report["eval"] = {
             "ppl": ppl,
-            "ppl_bf16_baseline_ref": 8.756,
-            "ppl_hsaq_baseline_ref": 10.013,
-            "pct_above_bf16": (ppl - 8.756) / 8.756 * 100,
-            "pct_above_hsaq": (ppl - 10.013) / 10.013 * 100,
+            "n_windows": n_windows,
+            "ppl_bf16_baseline_ref": 6.5482,
+            "ppl_hsaq_baseline_ref": 7.3814,
+            "pct_above_bf16": (ppl - 6.5482) / 6.5482 * 100,
+            "pct_above_hsaq": (ppl - 7.3814) / 7.3814 * 100,
         }
         report["status"] = "success"
         logger.info("PPL: %.4f (%.2f%% above bf16, %.2f%% vs HSAQ baseline)",