Fix DeepSeek V3.2 on Apple Silicon: Force CPU fallback for Sparse Top… #24
Conversation
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
…-K, disable fused decode, and fix model loader
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Outdated suggestions cannot be applied.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.
Fix DeepSeek V3.2 Inference on Apple Silicon (Metal)
Description:
This PR addresses multiple runtime crashes and assertions encountered when running the experimental DeepSeek V3.2 model on Apple Silicon (M3 Ultra).
Changes:
llama-model-loader.cpp: Fixed a crash during tensor loading where tensor->src[0] was being accessed before initialization. Rewrote the loop to correctly handle tensor naming and allocation.
llama-sparse-mla-fwd.cpp: Disabled LLAMA_SPARSE_MLA_FUSED_DECODE by default. The fused decode kernel is not currently implemented for Metal, causing assertions.
llama-sparse-topk.cpp: Added a fallback for SPARSE_TOPK_RADIX.
Issue: The Metal backend does not implement the SPARSE_TOPK_RADIX operator, causing the graph to fall back to CPU. However, the CPU backend explicitly asserts false for this operator, leading to a crash.
Fix: Wrapped the call in an #ifdef APPLE block to force the use of the CPU-based sparse_attn_topk::topk_radix_indices function instead of the graph operator, bypassing the missing backend support while preserving CUDA behavior for other platforms.
Result:
Model loads and runs inference successfully on M3 Ultra with ~12 t/s prompt processing and ~8 t/s generation.
…-K, disable fused decode, and fix model loader
Make sure to read the contributing guidelines before submitting a PR