hf2q 0.1.3

Pure Rust CLI for converting HuggingFace models to hardware-optimized formats and serving them over an OpenAI-compatible API on Apple Silicon
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165

#!/usr/bin/env python3
"""
Analyze whether Q8_0 quantization systematically lifts the <pad> (id 0)
logit relative to F16 quantization, by measuring:

  1. Per-row reconstruction error for pad vs median vs common tokens.
  2. Per-block amax distribution in pad row vs other rows (Q8_0 uses
     per-block amax/127 scale; near-zero blocks have compressed dynamic
     range).
  3. Empirical logit gap `Q8_logit - F16_logit` for pad on a distribution
     of realistic hidden states.

Inputs:
  - Gemma-4 GGUF model (embed_weight, F16 stored as F16, dequantized to F32).
  - Dumped hidden state from an actual decode run (pre-lm_head).

Output: numerical evidence for whether pad is a row-specific pathology
or just a near-tie manifestation.
"""

import gguf
import numpy as np
from pathlib import Path

GGUF_PATH = '/opt/hf2q/models/gemma-4-26B-A4B-it-ara-abliterated-dwq/gemma-4-26B-A4B-it-ara-abliterated-dwq.gguf'
HS = 2816
VOCAB = 262144

def quantize_q8_0_row(row_f32: np.ndarray) -> np.ndarray:
    """Q8_0 quantize a row, then dequantize back — returns the effective
    weight as the hf2q kernel would see it (F16 scale, int8 quants,
    scale = amax/127 per 32-block)."""
    assert row_f32.size % 32 == 0
    n_blocks = row_f32.size // 32
    out = np.zeros_like(row_f32)
    for b in range(n_blocks):
        block = row_f32[b*32:(b+1)*32]
        amax = float(np.abs(block).max())
        d_f32 = amax / 127.0
        # hf2q stores d as F16
        d_f16 = np.float32(np.float16(d_f32))
        if d_f16 == 0.0:
            inv_d = 0.0
        else:
            inv_d = 1.0 / d_f16
        q = np.clip(np.round(block * inv_d), -127, 127).astype(np.int8)
        out[b*32:(b+1)*32] = q.astype(np.float32) * d_f16
    return out

def quantize_f16_row(row_f32: np.ndarray) -> np.ndarray:
    """F16 roundtrip."""
    return np.float16(row_f32).astype(np.float32)

def rel_err(a: np.ndarray, b: np.ndarray) -> float:
    return float(np.sqrt(np.mean((a-b)**2)) / (np.sqrt(np.mean(a*a)) + 1e-12))

# --- Load embed ---
print("Loading GGUF embed_weight ...")
r = gguf.GGUFReader(GGUF_PATH)
embed = None
for t in r.tensors:
    if t.name == 'token_embd.weight':
        embed = np.array(t.data).astype(np.float32).reshape(VOCAB, HS)
        print(f"  shape={embed.shape} dtype={t.tensor_type.name}")
        break
assert embed is not None

# --- Per-row reconstruction error ---
print("\n=== Per-row reconstruction error (F16 vs Q8_0 vs F32 reference) ===")

# Sample tokens: pad, eos, bos, turn tokens, and a random sample
sample_ids = [0, 1, 2, 105, 106] + [int(i) for i in np.random.default_rng(42).integers(1000, VOCAB, 20)]

def row_stats(rid):
    row = embed[rid]
    f16 = quantize_f16_row(row)
    q8 = quantize_q8_0_row(row)
    return {
        'std': row.std(),
        'amax': np.abs(row).max(),
        'f16_err': rel_err(row, f16),
        'q8_err':  rel_err(row, q8),
        'f16_abs_err': float(np.abs(row - f16).mean()),
        'q8_abs_err':  float(np.abs(row - q8).mean()),
        'f16_max_abs': float(np.abs(row - f16).max()),
        'q8_max_abs':  float(np.abs(row - q8).max()),
    }

stats_by_id = {rid: row_stats(rid) for rid in sample_ids}
print(f"{'id':>8} {'std':>8} {'amax':>8} {'f16_rrms':>10} {'q8_rrms':>10} "
      f"{'f16_mae':>10} {'q8_mae':>10} {'q8/f16':>8}")
for rid in sample_ids:
    s = stats_by_id[rid]
    ratio = s['q8_err'] / s['f16_err'] if s['f16_err'] > 0 else float('inf')
    marker = "  ← pad" if rid == 0 else ""
    print(f"{rid:>8} {s['std']:>8.4f} {s['amax']:>8.4f} "
          f"{s['f16_err']:>10.3e} {s['q8_err']:>10.3e} "
          f"{s['f16_abs_err']:>10.3e} {s['q8_abs_err']:>10.3e} "
          f"{ratio:>8.2f}{marker}")

# --- Per-block amax distribution for pad row ---
print("\n=== Per-block amax distribution (88 blocks per row) ===")
pad_row = embed[0]
tok1k_row = embed[1000]  # random "typical" token

def per_block_amax(row):
    return np.array([np.abs(row[b*32:(b+1)*32]).max() for b in range(HS // 32)])

pad_amax = per_block_amax(pad_row)
tok_amax = per_block_amax(tok1k_row)
print(f"  pad block amax:  min={pad_amax.min():.4e}  max={pad_amax.max():.4e}  "
      f"mean={pad_amax.mean():.4e}  median={np.median(pad_amax):.4e}")
print(f"  tok1k amax:      min={tok_amax.min():.4e}  max={tok_amax.max():.4e}  "
      f"mean={tok_amax.mean():.4e}  median={np.median(tok_amax):.4e}")

# Number of "tiny" blocks (amax < 0.01) — these get amplified Q8 error
print(f"  pad blocks with amax < 0.01: {(pad_amax < 0.01).sum()}/{len(pad_amax)}")
print(f"  tok1k blocks with amax < 0.01: {(tok_amax < 0.01).sum()}/{len(tok_amax)}")

# --- Empirical F16 vs Q8 logit gap for pad on realistic hidden states ---
print("\n=== F16 vs Q8 logit for pad, across many hidden states ===")
# Use a distribution that mimics post-final-norm hidden state.
# From measurement: pre-lm_head has ~std=0.34 in this model (seen in earlier
# bisection). Generate N random hidden states.
rng = np.random.default_rng(0)
N = 512
H = rng.standard_normal((N, HS)).astype(np.float32) * 0.34

# Compute logits for rows: pad, top-10 typical
test_ids = [0, 1, 2, 105, 106, 100, 1000, 10000, 50000, 100000, 200000]

# Store full F16 and Q8 rows
rows_f16 = {rid: quantize_f16_row(embed[rid]) for rid in test_ids}
rows_q8 = {rid: quantize_q8_0_row(embed[rid]) for rid in test_ids}

def logits(h, row):
    return h @ row  # scalar: dot product

# For each hidden state, compute pad's F16 and Q8 logit, plus a reference logit (say tok 1000)
stats = {rid: {'gaps': []} for rid in test_ids}
for i in range(N):
    h = H[i]
    for rid in test_ids:
        l_f16 = float(h @ rows_f16[rid])
        l_q8 = float(h @ rows_q8[rid])
        stats[rid]['gaps'].append(l_q8 - l_f16)

print(f"  per-token (logit_Q8 - logit_F16) over {N} random hidden states (std=0.34):")
print(f"  {'id':>8} {'mean':>10} {'std':>10} {'min':>10} {'max':>10} {'max_abs':>10}")
for rid in test_ids:
    g = np.array(stats[rid]['gaps'])
    marker = "  ← pad" if rid == 0 else ""
    print(f"  {rid:>8} {g.mean():>10.4e} {g.std():>10.4e} {g.min():>10.4e} "
          f"{g.max():>10.4e} {np.abs(g).max():>10.4e}{marker}")

# --- What could make pad win? Check variance of pad's Q8 error projection on h ---
# logit_Q8[pad] - logit_F16[pad] = h @ (q8_pad - f16_pad)
# magnitude ~ ||h|| * ||q8_pad - f16_pad|| * cos(theta)
pad_f16 = rows_f16[0]
pad_q8 = rows_q8[0]
pad_err_vec = pad_q8 - pad_f16  # the "error direction" of Q8 vs F16 for pad
print(f"\n  pad Q8-F16 error vector: std={pad_err_vec.std():.4e}  norm={np.linalg.norm(pad_err_vec):.4e}")
# For tok1k
t1k_err = rows_q8[1000] - rows_f16[1000]
print(f"  tok1k Q8-F16 error vector: std={t1k_err.std():.4e}  norm={np.linalg.norm(t1k_err):.4e}")