Commit e3c3f666a for llama.cpp
commit e3c3f666a4ca1ca9994adea6a2b7f72a9746482a
Author: Pelayo Garcia Bárcena <40317024+PelaGB17@users.noreply.github.com>
Date: Fri Nov 20 22:52:19 2026 +0100
hip: add RDNA1 MMVQ parameter table (1 warp per block for ncols=1) (#32389)
* hip: add RDNA1 MMVQ parameter table (1 warp per block for ncols=1)
* hip: write the RDNA1 MMVQ values explicitly instead of recursing into the generic table
* hip: list the RDNA1 MMVQ table before RDNA2 in the enum
* Update ggml/src/ggml-cuda/mmvq.cu
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 091fc3071..6f10428e6 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -93,6 +93,7 @@ enum mmvq_parameter_table_id {
MMVQ_PARAMETERS_GENERIC = 0,
MMVQ_PARAMETERS_TURING,
MMVQ_PARAMETERS_GCN,
+ MMVQ_PARAMETERS_RDNA1,
MMVQ_PARAMETERS_RDNA2,
MMVQ_PARAMETERS_RDNA3_0,
MMVQ_PARAMETERS_RDNA4,
@@ -105,6 +106,8 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
return MMVQ_PARAMETERS_RDNA4;
#elif defined(RDNA3_0)
return MMVQ_PARAMETERS_RDNA3_0;
+#elif defined(RDNA1)
+ return MMVQ_PARAMETERS_RDNA1;
#elif defined(RDNA2) || defined(RDNA3_5) || defined(RDNA4_M)
return MMVQ_PARAMETERS_RDNA2;
#elif defined(GCN) || defined(CDNA)
@@ -127,6 +130,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
if (GGML_CUDA_CC_IS_RDNA3_0(cc)) {
return MMVQ_PARAMETERS_RDNA3_0;
}
+ if (GGML_CUDA_CC_IS_RDNA1(cc)) {
+ return MMVQ_PARAMETERS_RDNA1;
+ }
if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc)) {
return MMVQ_PARAMETERS_RDNA2;
}
@@ -538,6 +544,23 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
}
return 1;
}
+ if (table_id == MMVQ_PARAMETERS_RDNA1) {
+ switch (ncols_dst) {
+ case 1:
+ return 1;
+ case 2:
+ case 3:
+ case 4:
+ return 4;
+ case 5:
+ case 6:
+ case 7:
+ case 8:
+ return 2;
+ default:
+ return 1;
+ }
+ }
if (table_id == MMVQ_PARAMETERS_TURING) {
if (ncols_dst == 1) {
switch (type) {
@@ -607,7 +630,7 @@ static constexpr __host__ __device__ bool mmvq_two_rows_ncols1(ggml_type type) {
}
static constexpr __host__ __device__ int calc_rows_per_block(ggml_type type, int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
- if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING || table_id == MMVQ_PARAMETERS_GB10) {
+ if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING || table_id == MMVQ_PARAMETERS_GB10 || table_id == MMVQ_PARAMETERS_RDNA1) {
switch (ncols_dst) {
case 1:
if (small_k) {
@@ -616,7 +639,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(ggml_type type, int
if (table_id == MMVQ_PARAMETERS_GB10) {
return type == GGML_TYPE_IQ2_S || type == GGML_TYPE_IQ3_S ? 2 : 1;
}
- return table_id == MMVQ_PARAMETERS_GENERIC && mmvq_two_rows_ncols1(type) ? 2 : 1;
+ return (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_RDNA1) && mmvq_two_rows_ncols1(type) ? 2 : 1;
case 2:
case 3:
case 4: