| """Are some F32 tensors actually f16 bit-patterns shifted into the high halfword? |
| |
| blk.0.attn_kv_a_norm reads 0.00915 as f32, but its high 16 bits read as f16 give |
| 1.021 — and its low 16 bits are ZERO. output_norm reads a normal 1.21. If the |
| low-16-bits-zero pattern is ~100% for some tensors and ~0% for others, those tensors |
| were written with f16 values placed at bf16 positions, and every consumer of this |
| GGUF (including llama.cpp) sees weights ~110x too small. |
| |
| Decisive stat per tensor: fraction of elements whose low 16 bits are zero. |
| Random f32 data has essentially 0% (chance ~1/65536). |
| """ |
| import sys |
| import numpy as np |
| from gguf import GGUFReader |
|
|
| MODEL = r"E:\ABU\MODELS\llm\glm-5.2-ream\GLM-5.2-REAM-Q2_K.gguf" |
| reader = GGUFReader(MODEL, "r") |
|
|
| def stats(t): |
| n = int(np.prod([int(s) for s in t.shape])) |
| raw = np.ascontiguousarray(t.data).tobytes()[: n * 4] |
| u = np.frombuffer(raw, dtype=np.uint32) |
| f = np.frombuffer(raw, dtype=np.float32) |
| lo_zero = float((u & 0xFFFF == 0).mean()) |
| hi_as_f16 = np.frombuffer((u >> 16).astype(np.uint16).tobytes(), dtype=np.float16) |
| return n, lo_zero, float(np.sqrt((f * f).mean())), float(np.sqrt((hi_as_f16.astype(np.float32) ** 2).mean())) |
|
|
| TARGETS = [] |
| for L in (0, 12, 40, 77): |
| for nm in ("attn_norm", "attn_q_a_norm", "attn_kv_a_norm", "ffn_norm"): |
| TARGETS.append(f"blk.{L}.{nm}.weight") |
| TARGETS += ["output_norm.weight"] |
|
|
| by_name = {t.name: t for t in reader.tensors} |
| print(f"{'tensor':>32} {'n':>6} {'low16==0':>10} {'rms as f32':>12} {'rms hi-as-f16':>14} verdict") |
| print("-" * 104) |
| for nm in TARGETS: |
| t = by_name.get(nm) |
| if t is None or str(t.tensor_type) not in ("0", "GGMLQuantizationType.F32"): |
| continue |
| n, lz, r32, r16 = stats(t) |
| v = "F16-IN-HIGH-HALF" if lz > 0.98 else ("mixed" if lz > 0.02 else "genuine F32") |
| print(f"{nm:>32} {n:>6} {lz*100:>9.2f}% {r32:>12.5f} {r16:>14.5f} {v}") |
|
|
| |
| print("\n=== WHOLE-FILE SCAN of F32 tensors ===") |
| tot = bad = 0 |
| worst = [] |
| for t in reader.tensors: |
| if str(t.tensor_type) not in ("0", "GGMLQuantizationType.F32"): |
| continue |
| n, lz, r32, r16 = stats(t) |
| tot += 1 |
| if lz > 0.98: |
| bad += 1 |
| worst.append((t.name, r32, r16)) |
| print(f" F32 tensors: {tot} with low-16-bits all zero: {bad} ({100*bad/max(tot,1):.1f}%)") |
| fams = {} |
| for nm, r32, r16 in worst: |
| key = nm.split(".")[-2] if nm.startswith("blk.") else nm |
| fams.setdefault(key, []).append((r32, r16)) |
| for k, v in sorted(fams.items()): |
| a = np.mean([x[0] for x in v]); b = np.mean([x[1] for x in v]) |
| print(f" {k:>22} x{len(v):<4} mean rms f32={a:.5f} as-f16={b:.5f}" |
| f" ratio {b/max(a,1e-12):.1f}x") |
|
|
