Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-10 06:49:12 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: add TQ1_0 support (mm, mat-vec, mat-vec-id, dequant, get_rows) (#27765)

Browse Source 
* vulkan: add TQ1_0 support (mm, mat-vec, dequant, get_rows)

* vulkan: pack TQ1_0 powers of 3 into a 32-bit constant

Replaces the constant array with a packed 32-bit value (7 bits per entry,
max 81 < 128) extracted with shift/mask, as suggested in review — avoids a
constant array that may not be kept in registers.

test-backend-ops on gfx1151: tq1_0 MUL_MAT 11/11, MUL_MAT_ID 6/6,
GET_ROWS 4/4, unchanged.

* vulkan: address review - shared TQ1_0 decode helpers, fix standalone dequant shader

Review feedback from jeffbolznv, all points:

- Move the packed-pow3 decode into shared helpers in types.glsl
  (tq1_0_byte_of / tq1_0_digit_of / tq1_0_trit) and use them from
  dequant_funcs.glsl, mul_mm_funcs.glsl, dequant_funcs_cm2.glsl and
  dequant_tq1_0.comp instead of repeating the logic. The cm2 path also
  drops its constant array for the packed-constant extraction.
- Translate all remaining comments to English.
- dequant_tq1_0.comp: use dequant_head.glsl. The shader previously declared
  its own single-field push constant while the pipeline is created with the
  5-field layout, so p.ne read the wrong field - confirmed broken, as
  suspected in review.
- Fix wg_denoms for the standalone dequant pipeline: one invocation decodes
  4 elements with local_size 256, so a workgroup covers 256*4 elements, not
  256*16. With the old value the dispatcher launched a quarter of the
  required workgroups.

Verified by temporarily forcing the dequant + f16 matmul path for TQ1_0
(hack not committed): test-backend-ops MUL_MAT passes through the rewritten
standalone shader, and the standard MUL_MAT / MUL_MAT_ID / GET_ROWS
tq1_0 cases still pass on Vulkan (AMD gfx1151).

* vulkan: address review — English comments, shared tq1_0_trit, trim TQ1_0 test cases

- mul_mat_vec_tq1_0.comp: drop leftover non-English comment and the local
  POW3_PACKED constant; all decode sites now call tq1_0_trit() from types.glsl
- types.glsl / dequant_funcs_cm2.glsl: ASCII-only, drop stale reviewer note
- test-backend-ops: remove the oversized MUL_MAT_ID case (432 MiB A tensor,
  ~172 GFLOP reference); move the two remaining ones next to the other
  backend-specific mul_mat_id one-offs and document why they are needed

* metal: decline TQ1_0 for GET_ROWS and mat-mul in supports_op

The new TQ1_0 cases in test-backend-ops exposed that the Metal backend
claimed support for GET_ROWS/MUL_MAT/MUL_MAT_ID with TQ1_0 sources while
having no such kernels (ggml_metal_library_compile_pipeline aborted on the
missing kernel_get_rows_tq1_0). Decline the type so the ops fall back to
the CPU, matching the existing NVFP4 handling on the same lines.

Assisted-by: Claude Fable 5

* vulkan: trim the TQ1_0 comments

Addresses @0cc4m's review: keep only what the code does not already say.

Removed the block-format recaps (the layout is right there in the struct) and
the step-by-step decode walkthrough. Kept the two facts a reader cannot infer:
the 8-bit truncation is part of the format, not an optimisation, and the powers
of 3 are packed into one uint so they do not end up in a constant array that
may miss the registers.

No functional change.

* vulkan: address review — trim comments, fold Metal check, drop unused _v

Per @0cc4m's review:

- dequant_funcs.glsl, dequant_funcs_cm2.glsl: drop the "see types.glsl"
  pointers — they apply to every quant and say nothing specific.
- dequant_tq1_0.comp: drop the wg_denoms note. It is a precondition, not
  information.
- mul_mm_funcs.glsl: same pointer removed.
- types.glsl: the comment on tq1_0_trit is down to the one fact the code
  cannot show — the 8-bit truncation is part of the format, matching the C
  reference, not an optimisation.
- dequant_funcs_cm2.glsl: removed dequantFuncTQ1_0_v and its define. You were
  right that it is optional: it wrapped four scalar decodes and vectorised
  nothing, and mul_mm_cm2.comp already guards the path with
  `#if defined(dequantFuncA_v)` (DATA_A_F32 omits it the same way).
- ggml-metal-device.m: folded TQ1_0 into the existing NVFP4 check instead of a
  separate block, and dropped both comments.
- test-backend-ops.cpp: the two mul_mat_id cases stay — they cover the
  block-stride loop and the per-expert base offset that k == 256 alone never
  reaches — but the comment is now one line instead of five.

Kept: the one-line labels on the three block regions in mul_mat_vec_tq1_0.comp
and on tq1_0_byte_of(). Those state the 5-trits-per-byte packing, which the
loop bounds do not show. Happy to remove them too if you prefer.

Re-verified on AMD gfx1151 (Vulkan), test-backend-ops, 2/2 backends passed:
MUL_MAT 9 TQ1_0 cases, MUL_MAT_ID 5, GET_ROWS 4 — all OK, no failures.
The coopmat2 path is unchanged apart from the removed _v define.
This commit is contained in:
Anjielon
2026-09-07 06:35:30 +02:00
committed by GitHub
parent 465e49b9ce
commit 8fe90e1fbf
10 changed files with 233 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-device.m
+5 -2
View File
@@ -1486,7 +1486,9 @@ static bool ggml_metal_supports_mul_mat_op(
const struct ggml_tensor * op,
bool src0_f16_has_mv,
bool mm_path) {
if (!has_simdgroup_reduction || op->src[0]->type == GGML_TYPE_NVFP4) {
if (!has_simdgroup_reduction ||
op->src[0]->type == GGML_TYPE_NVFP4 ||
op->src[0]->type == GGML_TYPE_TQ1_0) {
return false;
}
@@ -1887,7 +1889,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
};
}
case GGML_OP_GET_ROWS:
return op->src[0]->type != GGML_TYPE_NVFP4;
return op->src[0]->type != GGML_TYPE_NVFP4 &&
op->src[0]->type != GGML_TYPE_TQ1_0;
case GGML_OP_SET_ROWS:
{
if (op->src[0]->type == GGML_TYPE_F16) {
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+23
View File
@@ -4727,6 +4727,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0], matmul_q8_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K], matmul_q2_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_TQ2_0], matmul_tq2_0_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_TQ1_0], matmul_tq1_0_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K], matmul_q3_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K], matmul_q4_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K], matmul_q5_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
@@ -4768,6 +4769,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0], matmul_id_subgroup_tq1_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
@@ -4841,6 +4843,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0], matmul_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
CREATE_MM2(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ1_0], matmul_tq1_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -4886,6 +4889,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0], matmul_id_subgroup_tq1_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
@@ -4977,6 +4981,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0], matmul_tq2_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ1_0], matmul_tq1_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -5026,6 +5031,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_subgroup_tq2_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0], matmul_id_subgroup_tq1_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -5074,6 +5080,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0], matmul_id_tq2_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0], matmul_id_tq1_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -5154,6 +5161,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ2_0].f32acc, matmul_tq2_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_TQ1_0].f32acc, matmul_tq1_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -5202,6 +5210,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0].f32acc, matmul_id_subgroup_tq2_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0].f32acc, matmul_id_subgroup_tq1_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -5232,6 +5241,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_TQ2_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ2_0].f32acc, matmul_id_tq2_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_TQ1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_TQ1_0].f32acc, matmul_id_tq1_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -5341,6 +5351,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_TQ2_0][i], "mul_mat_vec_tq2_0_f32_f32", arr_dmmv_tq2_0_f32_f32_len[reduc16], arr_dmmv_tq2_0_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_TQ1_0][i], "mul_mat_vec_tq1_0_f32_f32", arr_dmmv_tq1_0_f32_f32_len[reduc16], arr_dmmv_tq1_0_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5369,6 +5380,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_TQ2_0][i], "mul_mat_vec_tq2_0_f16_f32", arr_dmmv_tq2_0_f16_f32_len[reduc16], arr_dmmv_tq2_0_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_TQ1_0][i], "mul_mat_vec_tq1_0_f16_f32", arr_dmmv_tq1_0_f16_f32_len[reduc16], arr_dmmv_tq1_0_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5424,6 +5436,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", arr_dmmv_id_q8_0_f32_f32_len[reduc], arr_dmmv_id_q8_0_f32_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", arr_dmmv_id_q2_k_f32_f32_len[reduc16], arr_dmmv_id_q2_k_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_TQ2_0], "mul_mat_vec_id_tq2_0_f32", arr_dmmv_id_tq2_0_f32_f32_len[reduc16], arr_dmmv_id_tq2_0_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_TQ1_0], "mul_mat_vec_id_tq1_0_f32", arr_dmmv_id_tq1_0_f32_f32_len[reduc16], arr_dmmv_id_tq1_0_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", arr_dmmv_id_q3_k_f32_f32_len[reduc16], arr_dmmv_id_q3_k_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", arr_dmmv_id_q4_k_f32_f32_len[reduc16], arr_dmmv_id_q4_k_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", arr_dmmv_id_q5_k_f32_f32_len[reduc16], arr_dmmv_id_q5_k_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
@@ -5490,6 +5503,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_transpose[GGML_TYPE_Q8_0], "dequant_q8_0_transpose", dequant_q8_0_transpose_len, dequant_q8_0_transpose_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_k", dequant_q2_k_len, dequant_q2_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_TQ2_0], "dequant_tq2_0", dequant_tq2_0_len, dequant_tq2_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_TQ1_0], "dequant_tq1_0", dequant_tq1_0_len, dequant_tq1_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_k", dequant_q3_k_len, dequant_q3_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
@@ -5519,6 +5533,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q2_K], "get_rows_q2_k", get_rows_q2_k_len, get_rows_q2_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_TQ2_0], "get_rows_tq2_0", get_rows_tq2_0_len, get_rows_tq2_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_TQ1_0], "get_rows_tq1_0", get_rows_tq1_0_len, get_rows_tq1_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q3_K], "get_rows_q3_k", get_rows_q3_k_len, get_rows_q3_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_K], "get_rows_q4_k", get_rows_q4_k_len, get_rows_q4_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_K], "get_rows_q5_k", get_rows_q5_k_len, get_rows_q5_k_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -5548,6 +5563,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q2_K], "get_rows_q2_k_f32", get_rows_q2_k_f32_len, get_rows_q2_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_TQ2_0], "get_rows_tq2_0_f32", get_rows_tq2_0_f32_len, get_rows_tq2_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_TQ1_0], "get_rows_tq1_0_f32", get_rows_tq1_0_f32_len, get_rows_tq1_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q3_K], "get_rows_q3_k_f32", get_rows_q3_k_f32_len, get_rows_q3_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_K], "get_rows_q4_k_f32", get_rows_q4_k_f32_len, get_rows_q4_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_K], "get_rows_q5_k_f32", get_rows_q5_k_f32_len, get_rows_q5_k_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -7787,6 +7803,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return nullptr;
@@ -7862,6 +7879,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return nullptr;
@@ -7932,6 +7950,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return nullptr;
@@ -8026,6 +8045,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return nullptr;
@@ -8099,6 +8119,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return nullptr;
@@ -18673,6 +18694,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
break;
default:
return false;
@@ -18779,6 +18801,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_TQ1_0:
case GGML_TYPE_I32:
return true;
default:
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
+15
View File
@@ -608,6 +608,21 @@ vec2 get_dm(uint ib, uint a_offset) {
}
#endif
#if defined(DATA_A_TQ1_0)
float tq1_0_val(uint ib, uint e, uint a_offset) {
const uint bidx = tq1_0_byte_of(e);
const uint qbyte = uint(bidx < 48u ? data_a[a_offset + ib].qs[bidx]
: data_a[a_offset + ib].qh[bidx - 48u]);
return float(tq1_0_trit(qbyte, tq1_0_digit_of(e))) - 1.0;
}
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(tq1_0_val(ib, iqs, a_offset), tq1_0_val(ib, iqs + 1u, a_offset));
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(float(data_a[a_offset + ib].d), 0);
}
#endif
#if defined(DATA_A_TQ2_0)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
// elem e -> byte qs[(e/128)*32 + e%32], bits 2*((e%128)/32); w = q - 1 (d applied via get_dm)
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
+15
View File
@@ -247,6 +247,19 @@ f16vec4 dequantFuncQ8_0_v(const in decodeBufQ8_0 bl, const in uint blockCoords[2
return f16vec4(vec4(qi) * vec4(float(d)));
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ1_0 {
block_tq1_0 block;
};
float16_t dequantFuncTQ1_0(const in decodeBufTQ1_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const uint e = coordInBlock[1];
const uint bidx = tq1_0_byte_of(e);
const uint qbyte = uint(bidx < 48u ? bl.block.qs[bidx] : bl.block.qh[bidx - 48u]);
const uint xi = tq1_0_trit(qbyte, tq1_0_digit_of(e));
return bl.block.d * (float16_t(int(xi)) - float16_t(1.0));
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ2_0 {
block_tq2_0 block;
};
@@ -1406,6 +1419,8 @@ f16vec4 dequantFuncNVFP4_v(const in decodeBufNVFP4 bl, const in uint blockCoords
#elif defined(DATA_A_Q8_0)
#define dequantFuncA dequantFuncQ8_0
#define dequantFuncA_v dequantFuncQ8_0_v
#elif defined(DATA_A_TQ1_0)
#define dequantFuncA dequantFuncTQ1_0
#elif defined(DATA_A_TQ2_0)
#define dequantFuncA dequantFuncTQ2_0
#define dequantFuncA_v dequantFuncTQ2_0_v
ggml/src/ggml-vulkan/vulkan-shaders/dequant_tq1_0.comp
+28
View File
@@ -0,0 +1,28 @@
#version 450
#include "dequant_head.glsl"
layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_tq1_0 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_GlobalInvocationID.x * 4;
if (i >= p.nel) {
return;
}
const uint ib = i / QUANT_K_TQ1_0;
const float d = float(data_a[ib].d);
[[unroll]] for (uint j = 0; j < 4 && (i + j) < p.nel; ++j) {
const uint e = (i + j) % QUANT_K_TQ1_0;
const uint bidx = tq1_0_byte_of(e);
const uint qbyte = uint(bidx < 48u ? data_a[ib].qs[bidx]
: data_a[ib].qh[bidx - 48u]);
const uint xi = tq1_0_trit(qbyte, tq1_0_digit_of(e));
data_b[i + j] = D_TYPE(d * (float(xi) - 1.0f));
}
}
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_tq1_0.comp
+85
View File
@@ -0,0 +1,85 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#include "mul_mat_vec_base.glsl"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
// Walks the packed bytes directly (byte m, digit t) rather than via
// tq1_0_byte_of()/tq1_0_digit_of(): one byte per thread, expanded in place.
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint tid = gl_LocalInvocationID.x;
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
for (uint nrow = 0; nrow < num_rows; ++nrow) {
const uint ib0 = a_offset + (first_row + nrow) * num_blocks_per_row;
for (uint jcol = 0; jcol < NUM_COLS; ++jcol) {
const uint b_base = (jcol * p.batch_stride_b);
for (uint i = tid/8; i < num_blocks_per_row; i += gl_WorkGroupSize.x/8) {
const FLOAT_TYPE d = float(data_a[ib0 + i].d);
// First qs chunk: 32 bytes (5*32 elements)
[[unroll]] for (uint m = tid%8; m < 32; m += 8) {
const uint q_byte = uint(data_a[ib0 + i].qs[m]);
[[unroll]] for (uint t = 0; t < 5; ++t) {
const uint xi = tq1_0_trit(q_byte, t);
const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
const uint elem = t * 32u + m;
const uint b_idx = i * QUANT_K + elem;
temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
}
}
// Second qs chunk: 16 bytes (5*16 elements)
[[unroll]] for (uint m = tid%8; m < 16; m += 8) {
const uint q_byte = uint(data_a[ib0 + i].qs[32u + m]);
[[unroll]] for (uint t = 0; t < 5; ++t) {
const uint xi = tq1_0_trit(q_byte, t);
const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
const uint elem = 160u + t * 16u + m;
const uint b_idx = i * QUANT_K + elem;
temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
}
}
// qh bytes: 4 bytes (4*4 elements)
[[unroll]] for (uint j = tid%8; j < 4; j += 8) {
const uint qh_byte = uint(data_a[ib0 + i].qh[j]);
[[unroll]] for (uint t = 0; t < 4; ++t) {
const uint xi = tq1_0_trit(qh_byte, t);
const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
const uint elem = 240u + t * 4u + j;
const uint b_idx = i * QUANT_K + elem;
temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
}
}
}
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
+18
View File
@@ -197,6 +197,24 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const uint k_pair = row * LOAD_VEC_A / 2;
store_a(col, k_pair, FLOAT_TYPEV2(v.xy));
store_a(col, k_pair + 1, FLOAT_TYPEV2(v.zw));
#elif defined(DATA_A_TQ1_0)
const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = (idx % 128) * 2; // element 0,2,4..254
const float d = float(data_a[ib].d);
vec2 v;
for (uint kk = 0u; kk < 2u; ++kk) {
const uint e = iqs + kk;
const uint bidx = tq1_0_byte_of(e);
const uint qbyte = uint(bidx < 48u ? data_a[ib].qs[bidx]
: data_a[ib].qh[bidx - 48u]);
v[kk] = d * (float(tq1_0_trit(qbyte, tq1_0_digit_of(e))) - 1.0);
}
const uint k_pair = row * LOAD_VEC_A / 2;
store_a(col, k_pair, FLOAT_TYPEV2(v.xy));
#elif defined(DATA_A_TQ2_0)
const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
ggml/src/ggml-vulkan/vulkan-shaders/types.glsl
+35
View File
@@ -303,6 +303,41 @@ struct block_q2_K_packed32
#define DATA_A_QUANT_K
#endif
#define QUANT_K_TQ1_0 256
// TQ1_0: base-3 packed trits, 5 per byte in `qs` (48B) and 4 in `qh` (4B).
struct block_tq1_0
{
uint8_t qs[(QUANT_K_TQ1_0 - 4 * QUANT_K_TQ1_0 / 64) / 5];
uint8_t qh[QUANT_K_TQ1_0 / 64];
float16_t d;
};
// Element e in [0,255] -> its packed byte (0..47 qs, 48..51 qh) and digit.
uint tq1_0_byte_of(uint e) {
return e < 160u ? (e % 32u)
: e < 240u ? 32u + ((e - 160u) % 16u)
: 48u + ((e - 240u) % 4u);
}
uint tq1_0_digit_of(uint e) {
return e < 160u ? (e / 32u)
: e < 240u ? ((e - 160u) / 16u)
: ((e - 240u) / 4u);
}
// The 8-bit truncation below is part of the format, not an optimisation:
// the C reference does `uint8_t q = qs[..] * pow3[n]`.
uint tq1_0_trit(uint qbyte, uint t) {
const uint POW3_PACKED = (1u << 28) | (3u << 21) | (9u << 14) | (27u << 7) | 81u;
return ((((qbyte * ((POW3_PACKED >> (7u * (4u - t))) & 0x7Fu)) & 255u) * 3u) >> 8);
}
#if defined(DATA_A_TQ1_0)
#define QUANT_K QUANT_K_TQ1_0
#define QUANT_R 1
#define A_TYPE block_tq1_0
#define DATA_A_QUANT_K
#endif
#define QUANT_K_TQ2_0 256
// ternary (BitNet): 2-bit codes, w = (q - 1) * d; qs layout matches q2_K's
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+2 -1
View File
@@ -72,6 +72,7 @@ const std::vector<std::string> type_names = {
"iq4_nl",
"mxfp4",
"nvfp4",
"tq1_0",
"tq2_0",
"bf16",
};
@@ -734,7 +735,7 @@ void process_shaders() {
for (const auto& tname : type_names) {
// mul mat vec
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0" || tname == "tq1_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}));
tests/test-backend-ops.cpp
+7 -2
View File
@@ -8646,7 +8646,7 @@ static const ggml_type all_types[] = {
GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_TQ2_0,
// GGML_TYPE_TQ1_0, // TODO: implement for all backends
GGML_TYPE_TQ1_0,
GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
@@ -8674,7 +8674,7 @@ static const ggml_type other_types[] = {
GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_TQ2_0,
// GGML_TYPE_TQ1_0, // TODO: implement for all backends
GGML_TYPE_TQ1_0,
GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,
GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
@@ -9815,6 +9815,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_MXFP4, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_Q4_0, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
// multiple blocks per row: exercises the block-stride loop and the
// per-expert base offset, which k == 256 alone leaves untested
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_TQ1_0, GGML_TYPE_F32, 28, 10, false, 1024, 1, 4096));
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_TQ1_0, GGML_TYPE_F32, 128, 8, false, 1024, 1, 2048));
for (ggml_type type_a : all_types) {
test_cases.emplace_back(new test_mul_mat_id(type_a, GGML_TYPE_F32, 4, 2, false, 64, 16, 3*ggml_blck_size(type_a)));
}