mrhaoxx f36699affd feat(sft): AMX MoE SFT backend with LoRA support ... Complete SFT (Supervised Fine-Tuning) backend for MoE models using AMX SIMD: Core C++ implementation: - sft_moe.hpp: Forward/backward with LoRA fused operations (~5500 lines) - moe-sft-tp.hpp: Tensor-parallel wrapper for multi-NUMA - amx/moe-sft-tp.hpp: AMX-specific TP implementation - avx_kernels.hpp: AVX512 SIMD kernels for LoRA GEMM - amx_kernels.hpp: AMX tile kernels for Panel5 rank-outer optimization - worker_pool: RDTSC profiling, Chrome trace output, SFT timer infrastructure - ext_bindings.cpp: SFT MOE pybind bindings (BF16/INT8/INT4 + SkipLoRA variants) Python sft/ submodule (kt_kernel.sft): - base.py: BaseSFTMoEWrapper with buffer management (template method pattern) - amx.py: AMXSFTMoEWrapper (weight loading, C++ task construction) - autograd.py: KTMoEFunction (torch.autograd.Function for distributed training) - layer.py: KTMoELayerWrapper (nn.Module replacing HF MoE layers) - arch.py: MOEArchConfig (Qwen3/DeepSeek/Mixtral architecture detection) - weights.py: Expert weight extraction and checkpoint loading - lora.py: PEFT LoRA adaptation (view buffers, grad buffers, save/load adapter) - wrapper.py: wrap_moe_layers_with_kt_wrapper, load_kt_model, build_kt_device_map - config.py: KTConfig dataclass (DeepSpeed-style opaque config passthrough) - dist_utils.py: Distributed gather/scatter, checkpoint-phase detection Design decisions: - Rank-0-only expert pattern: only rank 0 holds C++ wrapper and expert weights - DeepSpeed-style integration: accelerate keeps only KTransformersPlugin (framework interaction fields), all logic in kt_kernel.sft - Inference isolation: importing kt_kernel does not load sft/ submodule - Old field name compatibility: _get_kt_config() converts kt_xxx→xxx automatically Verified: Qwen3-235B-A22B 4GPU AMXBF16 training, loss converges normally.		2026-04-08 23:11:00 +08:00
..
__init__.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
amx.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
arch.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
autograd.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
base.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
config.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
dist_utils.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
layer.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
lora.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
weights.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00
wrapper.py	feat(sft): AMX MoE SFT backend with LoRA support	2026-04-08 23:11:00 +08:00