feat: add auto-detection for Blackwell GPU architecture (sm_100a)#550
Open
yurekami wants to merge 2 commits intodeepseek-ai:mainfrom
Open
feat: add auto-detection for Blackwell GPU architecture (sm_100a)#550yurekami wants to merge 2 commits intodeepseek-ai:mainfrom
yurekami wants to merge 2 commits intodeepseek-ai:mainfrom
Conversation
Add automatic GPU architecture detection to properly set TORCH_CUDA_ARCH_LIST for Blackwell GPUs (sm_100a/10.0a). This fixes build and runtime errors when using DeepEP on Blackwell systems. Changes: - Add get_cuda_arch_from_device() function to detect GPU architecture - Auto-detect sm_100a for Blackwell, sm_90a for Hopper - Print detected architecture during build for user feedback - Fall back to 9.0 if detection fails Users can still override with TORCH_CUDA_ARCH_LIST environment variable. Fixes deepseek-ai#519 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…seek-ai#548) The low-latency P2P communication hangs when using 8-byte width token counts because the signaling buffer type (int) doesn't match the expected size. Changes: - Add LL_SIGNAL_BITS config (default 32) with ll_signal_t typedef - Add int64_t atomic primitives (ld_acquire_sys_global, st_release_sys_global) - Update buffer types from int* to ll_signal_t* for signaling buffers - Add conditional compilation for 64-bit NVSHMEM atomic operations - Fix boundary check to use sizeof(ll_signal_t) To enable 64-bit signaling: compile with -DLL_SIGNAL_BITS=64 Fixes: deepseek-ai#548 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
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
1 participant
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.
Summary
Add automatic GPU architecture detection to properly set
TORCH_CUDA_ARCH_LISTfor Blackwell GPUs (sm_100a/10.0a).Problem
When building DeepEP on Blackwell GPUs, users encounter errors like:
This is because the default
TORCH_CUDA_ARCH_LISTis set to9.0, which doesn't match Blackwell'ssm_100aarchitecture.Solution
get_cuda_arch_from_device()function to automatically detect GPU architecturesm_100afor Blackwell,sm_90afor Hopper9.0if detection fails (maintains backward compatibility)Users can still override with
TORCH_CUDA_ARCH_LISTenvironment variable.Test plan
Fixes #519
🤖 Generated with Claude Code