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

opencl: disable FA and MoE weights repack to work around compiler issues for Adreno 850 GPU (#25745)

Browse Source 
* opencl: workaround for A850 compiler compat

* opencl: fix DX compiler version parsing and cleanup

---------

Co-authored-by: Li He <lih@qti.qualcomm.com>
This commit is contained in:
Hongqiang Wang
2026-07-15 20:53:14 -07:00
committed by GitHub
co-authored by Li He
parent 602f828b4d
commit 3dafb585f8
3 changed files with 43 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl
+6 -4
View File
@@ -274,8 +274,9 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
shared_b[b_local_offset.y] = bx8_f16.hi;
// Dequantization
reg_a.lo = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.lo)) * s;
reg_a.hi = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.hi)) * s;
// Cast the e8m0 scale to half to satisfy E17 compilers
reg_a.lo = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.lo)) * (half)s;
reg_a.hi = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.hi)) * (half)s;
sub_group_barrier(CLK_LOCAL_MEM_FENCE);
@@ -304,8 +305,9 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
shared_b[b_local_offset.y] = bx8_f16.hi;
// Dequantization
reg_a.lo = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.lo)) * s;
reg_a.hi = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.hi)) * s;
// Cast the e8m0 scale to half to satisfy E17 compilers
reg_a.lo = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.lo)) * (half)s;
reg_a.hi = mxfp4_to_fp16_packed8(as_ushort2(mxfp4x16.hi)) * (half)s;
sub_group_barrier(CLK_LOCAL_MEM_FENCE);
ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl
+2
View File
@@ -1,3 +1,5 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#ifdef cl_intel_required_subgroup_size
#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
#define INTEL_GPU 1