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
//! Debug Q6_K layout for V weight
use realizar::gguf::{MappedGGUFModel, OwnedQKVWeights, OwnedQuantizedModel};
use realizar::quantize::dequantize_q6_k;
fn main() {
let path = "/tmp/parity-bench/tinyllama-1.1b-q4_k_m.gguf";
let mapped = MappedGGUFModel::from_path(path).expect("Failed");
let model = OwnedQuantizedModel::from_mapped(&mapped).expect("test");
let layer = &model.layers()[0];
let OwnedQKVWeights::Separate {
q: _,
k: _,
v: v_weight,
} = &layer.qkv_weight
else {
panic!("Expected separate")
};
// V weight: in_dim=2048, out_dim=256
// Total elements: 2048 * 256 = 524,288
// Q6_K superblocks: 524,288 / 256 = 2048 superblocks
// Each superblock: 210 bytes, 256 values
println!(
"Weight shape: [in_dim={}, out_dim={}]",
v_weight.in_dim, v_weight.out_dim
);
println!("Total elements: {}", v_weight.in_dim * v_weight.out_dim);
println!(
"Num superblocks: {}",
(v_weight.in_dim * v_weight.out_dim) / 256
);
// Dequantize first superblock (first 256 values)
let sb0 = dequantize_q6_k(&v_weight.data[..210]).expect("test");
println!("\nSuperblock 0 (first 256 values):");
println!(" Values 0..5: {:?}", &sb0[..5]);
println!(" Values 251..256: {:?}", &sb0[251..256]);
// Dequantize second superblock (values 256..512)
let sb1 = dequantize_q6_k(&v_weight.data[210..420]).expect("test");
println!("\nSuperblock 1 (values 256..512):");
println!(" Values 0..5: {:?}", &sb1[..5]);
// Now dequantize all and check what the layout is
let num_blocks = 2048;
let mut all_vals = Vec::new();
for i in 0..num_blocks {
let block = dequantize_q6_k(&v_weight.data[i * 210..(i + 1) * 210]).expect("test");
all_vals.extend(block);
}
// HuggingFace V weight is [256, 2048] in row-major
// So W[0, :] is the first row (2048 values)
// And W[:, 0] is the first column (256 values)
// Let's check: is the dequantized data stored as:
// Option A: [256 rows, 2048 cols] row-major => first 2048 values = W[0, :]
// Option B: [2048 cols, 256 rows] col-major => first 256 values = W[:, 0]
// HuggingFace W[0, :5] = [0.0281, 0.0059, -0.0003, -0.0056, 0.0075]
// HuggingFace W[:5, 0] = [0.0281, 0.0176, 0.0359, 0.0165, -0.0222]
println!("\nChecking layout interpretation:");
println!("If row-major [256, 2048], W[0, :5] = {:?}", &all_vals[..5]);
println!(
"If col-major [2048, 256], W[:5, 0] = {:?}",
[
all_vals[0],
all_vals[2048],
all_vals[4096],
all_vals[6144],
all_vals[8192]
]
);
println!("\nHuggingFace reference:");
println!(" W[0, :5] = [0.0281, 0.0059, -0.0003, -0.0056, 0.0075]");
println!(" W[:5, 0] = [0.0281, 0.0176, 0.0359, 0.0165, -0.0222]");
// The fused_q6k_parallel_matvec assumes:
// - 2048 columns (in_dim)
// - Each column has one superblock of 256 values
// - So superblock 0 = column 0 of W^T = row 0 of W
// Wait, that doesn't add up. Let me think again...
// Actually, the fused function has in_dim=2048 columns, each with 256 rows.
// That's a [2048, 256] matrix in column-major = W^T stored column-major.
// To get W, we need W[i, j] = data[j][i] where data is W^T column-major.
// For the matvec y = W @ x:
// y[i] = sum_j W[i, j] * x[j] = sum_j data[j][i] * x[j]
// Which is what the fused function computes.
// So the fused function should be correct IF the data is [2048, 256] column-major.
// Let me verify: is superblock 0 the first column (256 values for column 0)?
println!("\nSuperblock 0 should be W^T[:, 0] = W[0, :256] (first 256 of row 0)");
println!(" But HF W is [256, 2048], so W[0, :256] doesn't make sense...");
// Oh! I think I see the confusion. The dimensions are:
// - HF: W is [256, 2048], y = W @ x gives y[256], x[2048]
// - GGUF stores: [2048, 256] = W^T
// For y = W @ x where W is [256, 2048]:
// y[i] = sum_j W[i, j] * x[j] for j in 0..2048
// If GGUF stores W^T as [2048, 256] column-major:
// W^T[:, i] = W[i, :] is stored contiguously
// So to compute y[i] = sum_j W[i, j] * x[j], we need to access W[i, j] = W^T[j, i]
// The fused function accesses data in the pattern:
// for col in 0..2048:
// column = dequantize(superblock[col]) # 256 values = W^T[:, col] = W[*, col]
// for row in 0..256:
// output[row] += column[row] * x[col]
//
// This computes: output[row] += W[row, col] * x[col]
// Which is exactly y[row] = sum_col W[row, col] * x[col]
// So the algorithm is correct, but maybe the data isn't laid out as expected?
// Let me verify by checking what W[0, 0] should be vs what we get
println!("\nW[0, 0] should be HF W[0, 0] = 0.0281");
println!("Superblock 0, element 0 = {:?}", sb0[0]);
// If superblock 0 is W[:, 0] (first column), then sb0[0] = W[0, 0]
// If superblock 0 is W[0, :256] (first 256 of first row), then sb0[0] = W[0, 0] too
// Let's check W[0, 1] and W[1, 0]:
println!("W[0, 1] should be 0.0059 (HF)");
println!("W[1, 0] should be 0.0176 (HF)");
println!("sb0[1] = {:?}", sb0[1]); // If col-major, this is W[1, 0]
// If row-major, this is W[0, 1]
}