[ROCm] Pair gfx1250 support with gfx1250-strict - #3639
Open
BLOrange-AMD wants to merge 1 commit into
Open
BLOrange-AMD wants to merge 1 commit into
BLOrange-AMD wants to merge 1 commit into
Conversation
gfx1250-strict is a distinct LLVM/HIP offload target with the same CDNA5 ISA as gfx1250. Runtime equality checks against "gfx1250" alone miss it, and CMAKE HIP_ARCHITECTURES / CK GEMM filtering must treat both names as the same unsupported-CK family. Add gfx1250-strict next to gfx1250 in arch lists, hipBLASLt/hipSparseLt gates, shared-memory and workspace sizing, device preprocessor checks, wheel PYTORCH_ROCM_ARCH, and tests. Keep using the existing amd-gfx1250 AOTriton image, whose regex already matches gfx1250-strict. Test Plan: - rg gfx1250 to confirm remaining references include gfx1250-strict where they are architecture identifiers (AOTriton image name excluded). - No local ROCm gfx1250-strict hardware in this environment; unit tests that key off gcnArchName were updated in lockstep with CUDAHooks. Authored with an AI assistant. Co-authored-by: Bo Li <BLOrange-AMD@users.noreply.github.com>
naromero77amd
left a comment
There was a problem hiding this comment.
Nice first steps.
Right now we have gfx1250-strict for ROCm >=7.14. Most likely this will have to ROCm >= 10.1 or even 10.2.
| gcn_name = torch.cuda.get_device_properties(0).gcnArchName | ||
| return 'gfx950' in gcn_name or ('gfx1250' in gcn_name and ROCM_VERSION >= (7, 14)) | ||
| return 'gfx950' in gcn_name or ( | ||
| ('gfx1250' in gcn_name or 'gfx1250-strict' in gcn_name) and ROCM_VERSION >= (7, 14) |
There was a problem hiding this comment.
nitpick, 'gfx1250' in gcn_name also works when gcn_name is 'gfx1250-strict'
pruthvistony
requested changes
Sep 14, 2026
| #endif | ||
| #if ROCM_VERSION >= 71400 | ||
| "gfx1250", | ||
| "gfx1250", "gfx1250-strict", |
Collaborator
There was a problem hiding this comment.
gfx1250-strict is not available in 7.14, will need a separate ROCm version check
| #endif | ||
| #if ROCM_VERSION >= 71400 | ||
| "gfx1250", | ||
| "gfx1250", "gfx1250-strict", |
Collaborator
There was a problem hiding this comment.
gfx1250-strict is not available in 7.14, will need a separate ROCm version check
| { | ||
| __atomic_signal_fence(__ATOMIC_SEQ_CST); | ||
| #ifdef __gfx1250__ | ||
| #if defined(__gfx1250__) || defined(__gfx1250_strict__) |
Collaborator
There was a problem hiding this comment.
Lets confirm if the compiler is defining the macro - __gfx1250_strict__
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
gfx1250-strict is a distinct LLVM/HIP offload target with the same CDNA5 ISA as gfx1250. Runtime equality checks against "gfx1250" alone miss it, and CMAKE HIP_ARCHITECTURES / CK GEMM filtering must treat both names as the same unsupported-CK family.
Add gfx1250-strict next to gfx1250 in arch lists, hipBLASLt/hipSparseLt gates, shared-memory and workspace sizing, device preprocessor checks, wheel PYTORCH_ROCM_ARCH, and tests. Keep using the existing amd-gfx1250 AOTriton image, whose regex already matches gfx1250-strict.
Test Plan: