Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

graph : remove redundant GDN state transposes (#20443)

Browse Source 
* ggml : transpose fused GDN state access for coalesced memory reads (#20436)

The fused Gated Delta Net kernel accessed the [S_v, S_v] state matrix
column-wise on row-major storage, causing strided reads (stride S_v =
128 floats = 512 bytes) that waste GPU cache bandwidth. This produced a
39% regression on Qwen3.5-9B (Metal, M4 Max) compared to the unfused
path.

Transpose the state indexing so threads read contiguously:
- Metal: s_ptr[is*S_v] -> s_ptr[is] (stride 1 vs S_v)
- CUDA:  curr_state[i*S_v+col] -> curr_state[col*S_v+i] (coalesced)
- CPU:   restructured loops for row-wise transposed access

Also add --fused-gdn [on|off|auto] CLI flag (mirrors --flash-attn) so
users can control fused GDN independently of auto-detection.

All GATED_DELTA_NET backend-ops tests pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* ggml : use SIMD dot products in CPU GDN kernel, couple AR/chunked fused flags

- Replace scalar inner loops with ggml_vec_dot_f32 for SIMD-optimized
  dot products in the CPU fused GDN kernel (delta and attention output)
- Couple fused_gdn_ar and fused_gdn_ch flags in auto-detection: if one
  path lacks device support, disable both to prevent state layout mismatch
  between transposed (fused) and non-transposed (unfused) formats

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* llama : rever fgdn argument changes

* graph : remove GDN state transposes

* vulkan : adapt

* cuda : remove obsolete smem code

---------

Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Oliver Simons <osimons@nvidia.com>
This commit is contained in:
Georgi Gerganov
2026-03-13 22:12:54 +02:00
committed by GitHub
co-authored by Claude Opus 4.6 Paul Flynn Oliver Simons
parent 1430c35948
commit e30f1fdf74
5 changed files with 46 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ops.cpp
+21 -15
View File
@@ -10477,34 +10477,40 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
const float beta_val = *(const float *)((const char *)src_beta->data + iv3 * nbb3 + t * nbb2 + iv1 * nbb1);
const float * g_d = (const float *)((const char *)src_g->data + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
// state is stored transposed: s_out[j*S_v + i] = S[i][j]
// so row j of s_out = column j of S (contiguous access)
if (kda) {
// precompute exp(g) into delta scratch (reused below)
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_scale_f32(S_v, &s_out[i * S_v], expf(g_d[i]));
delta[i] = expf(g_d[i]);
}
// S[i][:] *= exp(g[i]) => for each row j of M: M[j][i] *= exp(g[i])
for (int64_t j = 0; j < S_v; ++j) {
ggml_vec_mul_f32(S_v, &s_out[j * S_v], &s_out[j * S_v], delta);
}
} else {
ggml_vec_scale_f32(S_v * S_v, s_out, expf(g_d[0]));
}
// delta[j] = sum_i S[j][i] * k[i]
memset(delta, 0, S_v * sizeof(float));
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, delta, &s_out[i * S_v], k_d[i]);
}
// delta[j] = sum_i S[i][j] * k[i] = dot(row j of M, k)
for (int64_t j = 0; j < S_v; ++j) {
delta[j] = (v_d[j] - delta[j]) * beta_val;
float sum = 0.0f;
ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, k_d, 0, 1);
delta[j] = (v_d[j] - sum) * beta_val;
}
// outer product: S[j][i] += k[i] * delta[j]
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, &s_out[i * S_v], delta, k_d[i]);
// outer product: S[i][j] += k[i] * delta[j] => M[j][i] += delta[j] * k[i]
for (int64_t j = 0; j < S_v; ++j) {
ggml_vec_mad_f32(S_v, &s_out[j * S_v], k_d, delta[j]);
}
// attn_out[j] = sum_i S[j][i] * q[i]
memset(attn_data, 0, S_v * sizeof(float));
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, attn_data, &s_out[i * S_v], q_d[i]);
// attn_out[j] = sum_i S[i][j] * q[i] = dot(row j of M, q)
for (int64_t j = 0; j < S_v; ++j) {
float sum = 0.0f;
ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, q_d, 0, 1);
attn_data[j] = sum * scale;
}
ggml_vec_scale_f32(S_v, attn_data, scale);
attn_data += S_v * H; // advance to next token
}