[MFM-2025-02-03] Merge Main to llama fp8; With Faster ROCm Paged Attention#399
Merged
hongxiayang merged 752 commits intoROCm:llama_fp8_12062024from EmbeddedLLM:main-to-llama-fp8Feb 3, 2025
+17,085-9,053
Commits
This pull request is big! We're only showing the most recent 250 commits
Commits on Jan 12, 2025
- authored
- authored
- authored
- authored
- authored
Commits on Jan 13, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- committed
- committed
- committed
- authored
- committed
Commits on Jan 14, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- committed
- committed
- authored
- authored
- authored
Commits on Jan 15, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 16, 2025
[Core] Default to using per_token quantization for fp8 when cutlass is supported. (vllm-project#8651)
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 17, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
[V1] Move more control of kv cache initialization from model_executor to EngineCore (vllm-project#11960)
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 18, 2025
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 19, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 20, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- committed
- authored
- authored
- authored
- authored
Commits on Jan 21, 2025
- authored
[Kernel] optimize moe_align_block_size for cuda graph and large num_experts (e.g. DeepSeek-V3) (vllm-project#12222)
authored- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
[Documentation][AMD] Add information about prebuilt ROCm vLLM docker for perf validation purpose (vllm-project#12281)
authored
Commits on Jan 22, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 23, 2025
[AMD][Quantization] Add TritonScaledMMLinearKernel since int8 is broken for AMD (vllm-project#12282)
authored- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
[BugFix] Fix parameter names and
process_after_weight_loadingfor W4A16 MoE Group Act Order (vllm-project#11528)authored- authored
- authored
- authored
Commits on Jan 24, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
[Bugfix][Kernel] FA3 Fix - RuntimeError: This flash attention build only supports pack_gqa (for build size reasons). (vllm-project#12405)
authored
Commits on Jan 25, 2025
- authored
- authored
- authored
- authored
- authored
Commits on Jan 26, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
Commits on Jan 27, 2025
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- authored
- committed
