Skip to content

fix(gguf): shape.rs byte-size table covers the IQ/TQ ggml types (#3601) #3654

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
alfredodeza wants to merge 2 commits into from
Closed

fix(gguf): shape.rs byte-size table covers the IQ/TQ ggml types (#3601) #3654

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8b39f2a
fix(gguf): shape.rs byte-size table covers the IQ/TQ ggml types (#3601)
alfredodeza Sep 21, 2026
f60304b
Merge branch 'main' into fix/3601-shape-iq-byte-size
alfredodeza Sep 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 46 additions & 0 deletions crates/aprender-core/src/format/gguf/builder.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,6 +94,52 @@ fn test_from_bytes_with_one_f32_tensor() {
assert!((extracted[3] - 4.0).abs() < f32::EPSILON);
}

// ========================================================================
// #3601: get_tensor_raw byte-size table must cover every IQ/TQ ggml id,
// not just the K-quants — real Unsloth "UD-*" dynamic quants mix these
// into otherwise-K-quant GGUFs, and `apr inspect` needs a size for every
// tensor it lists, not just the ones it can dequantize.
// ========================================================================

#[test]
fn test_get_tensor_raw_sizes_every_iq_and_tq_id() {
// (dtype, blck_size, type_size, name) — mirrors aprender-serve's vetted
// GGML_TYPES (gguf/ggml_type_table.rs, #3432), which shape.rs's byte_size
// match must agree with.
let cases: [(u32, usize, usize, &str); 11] = [
(16, 256, 66, "IQ2_XXS"),
(17, 256, 74, "IQ2_XS"),
(18, 256, 98, "IQ3_XXS"),
(19, 256, 50, "IQ1_S"),
(20, 32, 18, "IQ4_NL"),
(21, 256, 110, "IQ3_S"),
(22, 256, 82, "IQ2_S"),
(23, 256, 136, "IQ4_XS"),
(29, 256, 56, "IQ1_M"),
(34, 256, 54, "TQ1_0"),
(35, 256, 66, "TQ2_0"),
];

for (dtype, blck_size, type_size, name) in cases {
// One full block's worth of elements, so byte_size is exact.
let tensor_data = vec![0u8; type_size];
let data = build_synthetic_gguf_with_tensor(
"t.weight",
&[blck_size as u64],
dtype,
&tensor_data,
&[],
);
let reader = GgufReader::from_bytes(data).expect("parse GGUF with tensor");
let (raw, shape, got_dtype) = reader
.get_tensor_raw("t.weight")
.unwrap_or_else(|e| panic!("{name} (dtype {dtype}) must be sized, not refused: {e}"));
assert_eq!(raw.len(), type_size, "{name} byte size");
assert_eq!(shape, vec![blck_size]);
assert_eq!(got_dtype, dtype);
}
}

#[test]
fn test_from_bytes_tensor_excessive_dims() {
// n_dims > MAX_DIMS should fail
Expand Down
14 changes: 14 additions & 0 deletions crates/aprender-core/src/format/gguf/shape.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,7 +234,21 @@ impl GgufReader {
13 => (num_elements / 256).checked_mul(176), // Q5_K: 256 elements = 176 bytes
14 => (num_elements / 256).checked_mul(210), // Q6_K: 256 elements = 210 bytes
15 => (num_elements / 256).checked_mul(292), // Q8_K: 256 elements = 292 bytes
// #3601: IQ-family (codebook/lattice quant) and ternary types, sized from
// aprender-serve's vetted `GGML_TYPES` (gguf/ggml_type_table.rs, #3432) —
// real Unsloth "UD-*" dynamic quants mix these into otherwise-K-quant files.
16 => (num_elements / 256).checked_mul(66), // IQ2_XXS
17 => (num_elements / 256).checked_mul(74), // IQ2_XS
18 => (num_elements / 256).checked_mul(98), // IQ3_XXS
19 => (num_elements / 256).checked_mul(50), // IQ1_S
20 => (num_elements / 32).checked_mul(18), // IQ4_NL
21 => (num_elements / 256).checked_mul(110), // IQ3_S
22 => (num_elements / 256).checked_mul(82), // IQ2_S
23 => (num_elements / 256).checked_mul(136), // IQ4_XS
29 => (num_elements / 256).checked_mul(56), // IQ1_M
30 => num_elements.checked_mul(2), // BF16: 2 bytes per element
34 => (num_elements / 256).checked_mul(54), // TQ1_0
35 => (num_elements / 256).checked_mul(66), // TQ2_0
_ => {
return Err(AprenderError::FormatError {
message: format!("Unsupported dtype {} for raw extraction", meta.dtype),
Expand Down
Loading