Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-18 16:55:05 +02:00
Code Issues Packages Projects Releases Wiki Activity

use float for scales

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 10:41:33 +02:00
parent 1d2c5da46c
commit 6ca9b4548a
2 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
+4 -4
View File
@@ -86,10 +86,10 @@ const uint QPITCH = BK_STEP * (BK / 4) + 4;
// Shared memory cache
shared uint32_t buf_a_qs[BM * QPITCH];
shared float16_t buf_a_d[BM * BK_STEP];
shared float buf_a_d[BM * BK_STEP];
shared uint32_t buf_b_qs[BN * QPITCH];
shared float16_t buf_b_d[BN * BK_STEP];
shared float buf_b_d[BN * BK_STEP];
#define LOAD_VEC_A (4 * QUANT_R)
#define LOAD_VEC_B 16
@@ -283,7 +283,7 @@ void main() {
const uint idx = li * BK_STEP + ks; \
STORE_A_QS(buf_ib, ks, idx) \
if (loadr_a == 0) { \
buf_a_d[ks * BM + buf_ib] = pre_a_d[idx]; \
buf_a_d[ks * BM + buf_ib] = float(pre_a_d[idx]); \
} \
} \
} \
@@ -301,7 +301,7 @@ void main() {
buf_b_qs[base + 2] = v.z; \
buf_b_qs[base + 3] = v.w; \
if (loadr_b == 0) { \
buf_b_d[ks * BN + buf_ib] = in_bounds ? pre_b_d[idx] : float16_t(0.0); \
buf_b_d[ks * BN + buf_ib] = in_bounds ? float(pre_b_d[idx]) : 0.0f; \
} \
} \
} \
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1_funcs.glsl
+3 -3
View File
@@ -29,7 +29,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const ui
if (iqs == 0) {
#ifdef DATA_A_Q4_0
buf_a_d[ks * BM + buf_ib] = FLOAT_TYPE(data_a_packed16[ib].d);
buf_a_d[ks * BM + buf_ib] = float(data_a_packed16[ib].d);
#else // DATA_A_Q4_1
#endif
}
@@ -51,7 +51,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const ui
buf_a_qs[buf_ib * QPITCH + ks * (BK / 4) + iqs] = vui;
if (iqs == 0) {
buf_a_d[ks * BM + buf_ib] = FLOAT_TYPE(data_a_packed16[ib].d);
buf_a_d[ks * BM + buf_ib] = float(data_a_packed16[ib].d);
}
}
#endif
@@ -83,7 +83,7 @@ void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const ui
const uint ib_inner = ib % 4;
if (iqs == 0) {
buf_b_d[ks * BN + buf_ib] = data_b[ib_outer].ds[ib_inner].x;
buf_b_d[ks * BN + buf_ib] = float(data_b[ib_outer].ds[ib_inner].x);
}
const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];