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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
// aprender#3975: `CudaExecutor::gemm` documents B as `[k, n]`, but for `m == 1` it
// swaps in the `Gemv` kernel, which reads B as `[n, k]`. The GPU/CPU trace test
// matched "the [k,n] buffer read as [n,k]" to 7 significant digits.
//
// Every `gemm` caller is pinned here at m = 1 AND m = 2 against a CPU oracle, on
// a NON-SQUARE weight, so a layout misread cannot hide and a fix that repairs one
// side of the m = 1 / m > 1 split cannot silently break the other:
//
// caller | passes B as | RED today at
// --------------------------------------------|----------------|-------------
// CudaExecutor::gemm (its own contract) | [k, n] | m = 1
// CudaScheduler::matmul (GpuModel) | [k, n] | m = 1
// CachedSync::batch_matmul_gpu_prefer_cuda | [out,in]=[n,k] | m = 2
// OwnedQuantizedModel::fused_matmul (cuda) | [out,in]=[n,k] | m = 2
//
// Fix (quorum: explicit layout at the API): `gemm` is `[k, n]` at every m;
// `gemm_bt` / `CudaScheduler::matmul_bt` take `[n, k]`. The fused_matmul CUDA
// branch was dead (nothing set `cuda_executor`) and was deleted; its row now
// pins that attaching an executor cannot route the product through `gemm` again.
//
// GPU tests: they SKIP (with a printed line) on a host without a CUDA device, so they
// live under `gguf::cuda::`, the module path ci.yml's `cuda-unit` lane (yoga, a real
// GPU) selects. That lane is where they are a gate.
#[cfg(all(test, feature = "cuda"))]
mod gemm_layout_tests_3975 {
const IN: usize = 4;
const OUT: usize = 3;
/// Row-major `W[out, in]` — the APR/GGUF weight contract.
fn w_out_in() -> Vec<f32> {
vec![
1.0, 0.0, -1.0, 2.0, //
0.5, 1.0, 0.0, -1.0, //
2.0, -2.0, 1.0, 0.0,
]
}
/// The same weight as `[in, out]` (= `[k, n]`), what `gemm` documents for B.
fn w_in_out() -> Vec<f32> {
let w = w_out_in();
(0..IN)
.flat_map(|i| (0..OUT).map(move |o| (i, o)))
.map(|(i, o)| w[o * IN + i])
.collect()
}
/// `m` input rows; row 0 alone is the m = 1 case.
fn x(m: usize) -> Vec<f32> {
[[1.0, 2.0, 3.0, 4.0], [-1.0, 0.5, 2.0, -3.0]][..m].concat()
}
/// CPU oracle: `y[r, o] = sum_i x[r, i] * W[o, i]`.
fn oracle(x: &[f32]) -> Vec<f32> {
let w = w_out_in();
x.chunks(IN)
.flat_map(|row| {
(0..OUT)
.map(|o| (0..IN).map(|i| row[i] * w[o * IN + i]).sum::<f32>())
.collect::<Vec<_>>()
})
.collect()
}
fn assert_matches(got: &[f32], m: usize, caller: &str) {
let want = oracle(&x(m));
let off = got.iter().zip(&want).any(|(g, w)| (g - w).abs() > 1e-4);
assert!(
got.len() == want.len() && !off,
"#3975 {caller} at m={m}: got {got:?}, CPU oracle y = x*W^T is {want:?} -- \
a mismatch here is the weight read in the wrong layout"
);
}
/// GPU-FREE, so it runs on every cuda-feature build: the new `GemmBtTiled`
/// module is not empty, declares the entry the launcher looks up, and ptxas
/// assembles it. (#3970 found a kernel whose generate_ptx returned "".)
#[test]
fn gemm_bt_tiled_ptx_names_its_entry_and_assembles() {
use crate::cuda::{CudaKernels, KernelType};
let kernels = CudaKernels::new();
let kt = KernelType::GemmBtTiled {
m: 37,
n: 45,
k: 70,
tile_size: 16,
};
let ptx = kernels.generate_ptx(&kt);
let name = kernels.kernel_name(&kt);
assert!(
ptx.contains(&format!(".visible .entry {name}(")),
"GemmBtTiled PTX ({} bytes) does not declare the entry `{name}` the launcher looks up",
ptx.len()
);
assert!(
ptx.is_ascii(),
"ptxas rejects non-ASCII anywhere in a module"
);
let declared = crate::test_ptxas::declared_target(&ptx);
let assemble =
|ptx: &str, first: &[&str]| crate::test_ptxas::assemble(ptx, "gemm_bt", first);
// Row 1: the real module assembles (at its declared target, or the first newer
// arch this ptxas still defines).
if let Err(e) = assemble(&ptx, &[declared.as_str()]) {
panic!("ptxas rejected GemmBtTiled: {e}");
}
// Row 2: the fallback engages. `sm_10` is defined by no ptxas, standing in for
// CI's yoga, whose CUDA no longer defines `sm_70` (run 35847926651).
let used = assemble(&ptx, &["sm_10"]).expect("the unknown-arch fallback must engage");
assert_ne!(used, "sm_10");
// Row 3: the fallback never masks a real error.
let bogus = ptx.replacen("ret;", "bogus.plant.u32 %r0, %r0;\n ret;", 1);
assert_ne!(bogus, ptx, "the plant must land");
assert!(
assemble(&bogus, &[declared.as_str()]).is_err(),
"a bogus instruction must fail ptxas"
);
}
#[test]
fn gemm_honours_its_k_n_contract_at_m1_and_m2() {
let mut exec = crate::cuda_executor_or_skip!(0);
for m in [1, 2] {
let mut c = vec![0.0f32; m * OUT];
exec.gemm(&x(m), &w_in_out(), &mut c, m as u32, OUT as u32, IN as u32)
.expect("gemm");
assert_matches(&c, m, "CudaExecutor::gemm");
}
}
#[test]
fn cuda_scheduler_matmul_honours_k_n_at_m1_and_m2() {
let mut sched = crate::cuda_scheduler_or_skip!();
for m in [1, 2] {
let got = sched
.matmul(&x(m), &w_in_out(), m, IN, OUT)
.expect("matmul");
assert_matches(&got, m, "CudaScheduler::matmul (GpuModel)");
}
}
/// One 4x3 fixture is an anecdote. Sweep shapes that straddle the 16- and
/// 32-wide tiles (and m = 1) through both entry points against the oracle.
#[test]
fn gemm_and_gemm_bt_agree_with_the_oracle_across_tile_boundaries() {
let mut exec = crate::cuda_executor_or_skip!(0);
for &(m, k, n) in &[
(1, 70, 45),
(2, 70, 45),
(17, 33, 16),
(37, 70, 45),
(64, 128, 96),
] {
let x: Vec<f32> = (0..m * k)
.map(|i| ((i * 7 % 13) as f32 - 6.0) * 0.25)
.collect();
let w: Vec<f32> = (0..n * k)
.map(|i| ((i * 5 % 11) as f32 - 5.0) * 0.125)
.collect(); // [n, k]
let w_kn: Vec<f32> = (0..k * n).map(|j| w[(j % n) * k + j / n]).collect();
let want: Vec<f32> = (0..m * n)
.map(|j| {
(0..k)
.map(|i| x[(j / n) * k + i] * w[(j % n) * k + i])
.sum()
})
.collect();
let (m32, n32, k32) = (m as u32, n as u32, k as u32);
let mut kn = vec![0.0f32; m * n];
exec.gemm(&x, &w_kn, &mut kn, m32, n32, k32).expect("gemm");
let mut nk = vec![0.0f32; m * n];
exec.gemm_bt(&x, &w, &mut nk, m32, n32, k32)
.expect("gemm_bt");
for (name, got) in [("gemm [k,n]", &kn), ("gemm_bt [n,k]", &nk)] {
let worst = got
.iter()
.zip(&want)
.map(|(g, w)| (g - w).abs())
.fold(0.0f32, f32::max);
assert!(
worst < 1e-3,
"#3975 {name} at (m,k,n)=({m},{k},{n}): max |err| {worst}"
);
}
}
}
fn test_model() -> crate::gguf::OwnedQuantizedModel {
use crate::gguf::{ArchConstraints, GGUFConfig};
let config = GGUFConfig {
architecture: "llama".to_string(),
constraints: ArchConstraints::from_architecture("llama"),
hidden_dim: 64,
intermediate_dim: 128,
num_layers: 1,
num_heads: 4,
num_kv_heads: 4,
vocab_size: 256,
context_length: 64,
rope_theta: 10000.0,
eps: 1e-5,
rope_type: 0,
explicit_head_dim: None,
query_pre_attn_scalar: None,
bos_token_id: None,
eos_token_id: None,
};
crate::api::test_helpers::create_test_quantized_model(&config)
}
#[test]
fn cached_batch_matmul_reads_a_dequantized_out_in_weight_at_m1_and_m2() {
let cached = crate::gguf::OwnedQuantizedModelCachedSync::new(test_model());
let has_cuda = cached
.get_cuda_scheduler()
.map(|g| g.is_some())
.unwrap_or(false);
if !has_cuda {
eprintln!("SKIP: no CudaScheduler -- the wgpu fallback is not what #3975 pins");
return;
}
for m in [1, 2] {
// The production callers (batch_ffn_gpu, batch_qkv_projection_gpu, ...)
// pass `dequantize_weight(..)` output: row-major [out, in].
let got = cached
.batch_matmul_gpu_prefer_cuda(&x(m), &w_out_in(), m, IN, OUT)
.expect("batch matmul");
assert_matches(&got, m, "CachedSync::batch_matmul_gpu_prefer_cuda");
}
}
#[test]
fn fused_matmul_cuda_dequant_fallback_reads_out_in_at_m1_and_m2() {
let exec = crate::cuda_executor_or_skip!(0);
let mut model = test_model();
model.cuda_executor = Some(std::sync::Mutex::new(exec));
// F32 has no native quantized GEMV: before #3975 every m took dequant +
// `gemm` here. That branch is deleted, so this must agree with the oracle.
let weight = crate::gguf::OwnedQuantizedTensor {
data: w_out_in().iter().flat_map(|v| v.to_le_bytes()).collect(),
in_dim: IN,
out_dim: OUT,
qtype: crate::gguf::GGUF_TYPE_F32,
};
for m in [1, 2] {
let got = model.fused_matmul(&x(m), &weight).expect("fused_matmul");
assert_matches(&got, m, "OwnedQuantizedModel::fused_matmul (cuda)");
}
}
}