mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

discover: Disable flash attention for Jetson Xavier (CC 7.2)

Browse Source 
GGML picks the wrong kernel and these systems fail with:
Sep 28 22:25:39 xavier ollama[48999]: //ml/backend/ggml/ggml/src/ggml-cuda/fattn-wmma-f16.cu:437:
ERROR: CUDA kernel flash_attn_ext_f16 has no device code compatible with CUDA arch 720. ggml-cuda.cu
was compiled for: __CUDA_ARCH_LIST__

Fixes #12442
This commit is contained in:
Jesse Gross
2025-10-07 11:37:58 -07:00
committed by Jesse Gross
parent 4e5d862ec4
commit aa45f7ce27
3 changed files with 17 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
discover/gpu.go
View File

@@ -2,7 +2,6 @@ package discover
import (
"context"
"fmt"
"log/slog"
"os"
"path/filepath"
@@ -62,17 +61,14 @@ func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
DependencyPath: dev.LibraryPath,
DriverMajor: dev.DriverMajor,
DriverMinor: dev.DriverMinor,
ComputeMajor: dev.ComputeMajor,
ComputeMinor: dev.ComputeMinor,
}
if dev.Library == "CUDA" || dev.Library == "ROCm" {
info.MinimumMemory = 457 * format.MebiByte
}
if dev.Library == "ROCm" {
info.Compute = fmt.Sprintf("gfx%x%02x", dev.ComputeMajor, dev.ComputeMinor)
if rocmDir != "" {
info.DependencyPath = append(info.DependencyPath, rocmDir)
}
} else {
info.Compute = fmt.Sprintf("%d.%d", dev.ComputeMajor, dev.ComputeMinor)
if dev.Library == "ROCm" && rocmDir != "" {
info.DependencyPath = append(info.DependencyPath, rocmDir)
}
resp = append(resp, info)
}

9
discover/types.go
View File

@@ -37,9 +37,10 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
UnreliableFreeMemory bool
// GPU information
filterID string // AMD Workaround: The numeric ID of the device used to filter out other devices
Name string `json:"name"` // user friendly name if available
Compute string `json:"compute"` // Compute Capability or gfx
filterID string // AMD Workaround: The numeric ID of the device used to filter out other devices
Name string `json:"name"` // user friendly name if available
ComputeMajor int `json:"compute_major"` // Compute Capability or gfx
ComputeMinor int `json:"compute_minor"`
// Driver Information - TODO no need to put this on each GPU
DriverMajor int `json:"driver_major,omitempty"`
@@ -173,7 +174,7 @@ func (l GpuInfoList) FlashAttentionSupported() bool {
for _, gpu := range l {
supportsFA := gpu.Library == "cpu" ||
gpu.Name == "Metal" || gpu.Library == "Metal" ||
(gpu.Library == "CUDA" && gpu.DriverMajor >= 7) ||
(gpu.Library == "CUDA" && gpu.DriverMajor >= 7 && !(gpu.ComputeMajor == 7 && gpu.ComputeMinor == 2)) || // We don't have kernels for Jetson Xavier
gpu.Library == "ROCm"
if !supportsFA {