Skip to content

cuda: add DS4_CUDA_MANAGED env var for full UMA pool access on Strix Halo #313

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kmc6042 wants to merge 1 commit into
base: rocm
Choose a base branch
from
+19 −5
Open

cuda: add DS4_CUDA_MANAGED env var for full UMA pool access on Strix Halo #313

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 18 additions & 5 deletions ds4_cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1302,9 +1302,22 @@ extern "C" ds4_gpu_tensor *ds4_gpu_tensor_alloc(uint64_t bytes) {
if (bytes == 0) bytes = 1;
ds4_gpu_tensor *t = (ds4_gpu_tensor *)calloc(1, sizeof(*t));
if (!t) return NULL;
if (!cuda_ok(cudaMalloc(&t->ptr, (size_t)bytes), "tensor alloc")) {
free(t);
return NULL;

if (getenv("DS4_CUDA_MANAGED") != NULL) {
/* Use cudaMallocManaged with cudaMemAttachGlobal so the allocation
* is GPU-accessible across all streams. On UMA platforms (Strix
* Halo, Grace-Hopper) this allocates from the full unified pool,
* bypassing the BIOS VRAM carve-out. */
if (!cuda_ok(cudaMallocManaged(&t->ptr, (size_t)bytes, cudaMemAttachGlobal),
"managed tensor alloc")) {
free(t);
return NULL;
}
} else {
if (!cuda_ok(cudaMalloc(&t->ptr, (size_t)bytes), "tensor alloc")) {
free(t);
return NULL;
}
}
t->bytes = bytes;
t->owner = 1;
Expand Down Expand Up @@ -6168,7 +6181,7 @@ extern "C" int ds4_gpu_attention_prefill_raw_heads_tensor(ds4_gpu_tensor *heads,
if (!tmp) return 0;
float *scores = tmp;
float *out_tmp = (float *)((char *)tmp + out_offset);
const float alpha = rsqrtf((float)head_dim);
const float alpha = 1.0f / sqrtf((float)head_dim);
const float beta = 0.0f;
cublasStatus_t st = cublasSgemmStridedBatched(g_cublas,
CUBLAS_OP_T,
Expand Down Expand Up @@ -6538,7 +6551,7 @@ static int attention_prefill_mixed_launch(
n_comp,
head_dim);
if (!cuda_ok(cudaGetLastError(), "attention mixed kv pack launch")) return 0;
const float alpha = rsqrtf((float)head_dim);
const float alpha = 1.0f / sqrtf((float)head_dim);
const float beta = 0.0f;
cublasStatus_t st = cublasSgemmStridedBatched(g_cublas,
CUBLAS_OP_T,
Expand Down
1 change: 1 addition & 0 deletions ds4_rocm.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@

#define cudaSuccess hipSuccess
#define cudaErrorNotSupported hipErrorNotSupported
#define cudaMemAttachGlobal hipMemAttachGlobal
#define cudaErrorInvalidValue hipErrorInvalidValue
#define cudaGetLastError hipGetLastError
#define cudaGetErrorString hipGetErrorString
Expand Down