happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: gate tiled GEMM and split-KV paths to preserve q8_0/q4_0 vec_dot semantics

This commit is contained in:
Tim Burke 2026-03-20 23:38:43 -04:00
parent 358bd71b52
commit 23e88631c4
1 changed files with 11 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cpu/ops.cpp
View File

@ -9171,8 +9171,12 @@ static void ggml_compute_forward_flash_attn_ext_f16(
const bool use_ref = params->use_ref;
// Split-KV: parallelize across KV chunks for single-query decode (token generation).
// Delegates to one_chunk which handles all supported types (F16, Q8_0, Q4_0, MXFP, etc).
// Only for types whose tiled/one_chunk paths produce identical results (f32, f16, MXFP).
// Standard quant types (q8_0, q4_0) must use the scalar path to preserve vec_dot semantics.
const bool kv_is_f32_f16_or_mxfp = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16
|| ggml_is_type_mxfp(k->type));
const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1)
&& kv_is_f32_f16_or_mxfp
&& q->type == GGML_TYPE_F32 && nek1 >= 512;
if (use_split_kv_path) {
@ -9230,8 +9234,14 @@ static void ggml_compute_forward_flash_attn_ext_f16(
const int64_t dr = (nr + nchunk - 1) / nchunk;
static constexpr int64_t Q_TILE_SZ = ggml_fa_tile_config::Q;
// Tiled GEMM path: dequant K/V to float, then simd_gemm.
// Only for types that natively dequant to float (f32, f16, MXFP).
// Standard quant types (q8_0, q4_0) must use the scalar one_chunk path
// to preserve vec_dot semantics and produce identical results to master.
bool use_tiled = !use_ref &&
(q->type == GGML_TYPE_F32 &&
kv_is_f32_f16_or_mxfp &&
(k->type == v->type || ggml_is_type_mxfp(k->type)) &&
neq1 >= Q_TILE_SZ);
#ifdef GGML_SIMD
use_tiled &= (DV % GGML_F32_EPR == 0);