mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-10-25 03:10:08 +08:00
yyj
e80614fc2d
* feat(sft): support distributed activation reuse policies * feat(sft): add frozen-base INT8 LoRA training * fix(sft): make INT8 expert LoRA rank-zero authoritative * fix(sft): preserve DeepSeek router LoRA routing * feat(sft): enable persistent INT8 LoRA training * perf(sft): accelerate INT8 VNNI with oneDNN BRGEMM * perf(int8): fuse oneDNN compensation into backward repack * [feat]: support BF16 expert LoRA training * [fix]: honor forwarded activation policy in SFT workers * feat(sft): add native block-FP8 routed expert LoRA * feat(sft): expose explicit expert placeholder ownership * fix(sft): publish fused adapter artifacts atomically * feat(sft): own artifact and adapter lifecycle contracts * fix(sft): harden artifact and rank-local contracts * fix(sft): auto-adapt owner before adapter restore * style(sft): keep lifecycle comments concise * fix(sft): require fused adapter manifests * test(sft): use spawn for distributed workers * fix(sft): preserve runtime checkpoint metadata * fix(sft): validate wrapped runtime configuration * fix(sft): preserve expert format provenance * fix(sft): own routed experts during device dispatch * test(sft): lock explicit quantization conflict * fix(cpu): make shared memory buffers lifetime-safe * release: prepare v0.7.0
187 lines
5.6 KiB
Python
187 lines
5.6 KiB
Python
# SFT (Supervised Fine-Tuning) submodule for kt-kernel
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
"""
|
|
SFT training support for KT-Kernel MoE.
|
|
|
|
This submodule adds training capabilities (forward/backward, LoRA, autograd,
|
|
distributed) on top of the inference-only kt_kernel base package.
|
|
|
|
Additional dependencies beyond base kt_kernel: torch.nn, torch.distributed, peft (optional).
|
|
"""
|
|
|
|
from .config import KTActivationPolicy, KTConfig
|
|
from .backend import (
|
|
FP8_BACKEND,
|
|
FP8_SFT_METHOD,
|
|
INT8_BACKEND,
|
|
INT8_SFT_METHOD,
|
|
get_fp8_runtime,
|
|
get_int8_runtime,
|
|
)
|
|
from .weight_manifest import validate_persistent_int8_weights
|
|
from .artifacts import (
|
|
FUSED_EXPERT_LORA_NAME,
|
|
KT_ADAPTER_MANIFEST_NAME,
|
|
KT_NON_EXPERT_MANIFEST_NAME,
|
|
KTAdapterManifest,
|
|
KTArtifactError,
|
|
KTNonExpertCache,
|
|
KTPretrainedLoadPlan,
|
|
claim_kt_routed_expert_subtrees,
|
|
hide_kt_int8_routed_experts_from_dispatch,
|
|
hide_kt_routed_experts_from_dispatch,
|
|
is_kt_int8_routed_expert_base_parameter,
|
|
is_kt_routed_expert_base_parameter,
|
|
is_kt_routed_expert_parameter_name,
|
|
is_kt_supported_moe_model,
|
|
load_kt_adapter_artifacts,
|
|
mark_kt_int8_routed_expert_base_parameters,
|
|
prepare_kt_int8_non_expert_device_map,
|
|
prepare_kt_non_expert_device_map,
|
|
project_kt_int8_routed_experts_out_of_device_map,
|
|
project_kt_routed_experts_out_of_device_map,
|
|
resolve_kt_pretrained_artifacts,
|
|
save_kt_adapter_artifacts,
|
|
validate_kt_prequantized_loading_info,
|
|
validate_kt_pretrained_load,
|
|
write_kt_non_expert_cache_manifest,
|
|
)
|
|
from .dist_utils import get_activation_checkpoint_context_fn
|
|
from .base import BaseSFTMoEWrapper, KExpertsSFTBuffer
|
|
from .amx import AMXSFTMoEWrapper
|
|
from .arch import (
|
|
MOEArchConfig,
|
|
get_moe_arch_config,
|
|
get_moe_module,
|
|
move_non_experts_to_gpu,
|
|
get_expert_device,
|
|
KTAMXError,
|
|
KTAMXNotAvailableError,
|
|
KTAMXModelNotSupportedError,
|
|
KTAMXConfigError,
|
|
)
|
|
from .autograd import KTMoEFunction
|
|
from .conv3d_compat import is_vlm_conv3d_compatible, patch_vlm_conv3d
|
|
from .layer import KTMoELayerWrapper
|
|
from .weights import (
|
|
extract_moe_weights,
|
|
get_kt_expert_placeholders,
|
|
load_block_fp8_experts_from_checkpoint_files,
|
|
load_experts_from_checkpoint_files,
|
|
load_experts_from_kt_weight_path,
|
|
INT8ExpertWeights,
|
|
BlockFP8ExpertWeights,
|
|
)
|
|
from .lora import (
|
|
KTAdaptationResult,
|
|
adapt_peft_lora,
|
|
kt_adapt_peft_lora,
|
|
get_kt_lora_named_params,
|
|
get_kt_lora_params,
|
|
get_kt_named_trainable_params,
|
|
get_kt_rank_local_parameter_names,
|
|
get_kt_trainable_params,
|
|
update_kt_lora_pointers,
|
|
sync_kt_lora_gradients,
|
|
save_lora_experts_to_adapter,
|
|
save_kt_moe_to_adapter,
|
|
load_lora_experts_from_adapter,
|
|
load_kt_moe_from_adapter,
|
|
LoRAExpertMLP,
|
|
LoRAExperts,
|
|
)
|
|
from .wrapper import (
|
|
wrap_moe_layers_with_kt_wrapper,
|
|
build_kt_device_map,
|
|
build_kt_device_map_simplified,
|
|
get_kt_loading_kwargs,
|
|
load_kt_model,
|
|
)
|
|
from .profiler import collect_kt_sft_profile, format_kt_sft_profile, reset_kt_sft_profile
|
|
|
|
__all__ = [
|
|
"KTConfig",
|
|
"KTActivationPolicy",
|
|
"INT8_BACKEND",
|
|
"INT8_SFT_METHOD",
|
|
"get_int8_runtime",
|
|
"FP8_BACKEND",
|
|
"FP8_SFT_METHOD",
|
|
"get_fp8_runtime",
|
|
"validate_persistent_int8_weights",
|
|
"FUSED_EXPERT_LORA_NAME",
|
|
"KT_ADAPTER_MANIFEST_NAME",
|
|
"KT_NON_EXPERT_MANIFEST_NAME",
|
|
"KTAdapterManifest",
|
|
"KTArtifactError",
|
|
"KTNonExpertCache",
|
|
"KTPretrainedLoadPlan",
|
|
"claim_kt_routed_expert_subtrees",
|
|
"hide_kt_int8_routed_experts_from_dispatch",
|
|
"hide_kt_routed_experts_from_dispatch",
|
|
"is_kt_int8_routed_expert_base_parameter",
|
|
"is_kt_routed_expert_base_parameter",
|
|
"is_kt_routed_expert_parameter_name",
|
|
"is_kt_supported_moe_model",
|
|
"load_kt_adapter_artifacts",
|
|
"mark_kt_int8_routed_expert_base_parameters",
|
|
"prepare_kt_int8_non_expert_device_map",
|
|
"prepare_kt_non_expert_device_map",
|
|
"project_kt_int8_routed_experts_out_of_device_map",
|
|
"project_kt_routed_experts_out_of_device_map",
|
|
"resolve_kt_pretrained_artifacts",
|
|
"save_kt_adapter_artifacts",
|
|
"validate_kt_prequantized_loading_info",
|
|
"validate_kt_pretrained_load",
|
|
"write_kt_non_expert_cache_manifest",
|
|
"get_activation_checkpoint_context_fn",
|
|
"BaseSFTMoEWrapper",
|
|
"KExpertsSFTBuffer",
|
|
"AMXSFTMoEWrapper",
|
|
"MOEArchConfig",
|
|
"get_moe_arch_config",
|
|
"get_moe_module",
|
|
"move_non_experts_to_gpu",
|
|
"get_expert_device",
|
|
"KTAMXError",
|
|
"KTAMXNotAvailableError",
|
|
"KTAMXModelNotSupportedError",
|
|
"KTAMXConfigError",
|
|
"KTMoEFunction",
|
|
"patch_vlm_conv3d",
|
|
"is_vlm_conv3d_compatible",
|
|
"KTMoELayerWrapper",
|
|
"extract_moe_weights",
|
|
"get_kt_expert_placeholders",
|
|
"load_block_fp8_experts_from_checkpoint_files",
|
|
"load_experts_from_checkpoint_files",
|
|
"load_experts_from_kt_weight_path",
|
|
"INT8ExpertWeights",
|
|
"BlockFP8ExpertWeights",
|
|
"KTAdaptationResult",
|
|
"adapt_peft_lora",
|
|
"kt_adapt_peft_lora",
|
|
"get_kt_lora_named_params",
|
|
"get_kt_lora_params",
|
|
"get_kt_named_trainable_params",
|
|
"get_kt_rank_local_parameter_names",
|
|
"get_kt_trainable_params",
|
|
"update_kt_lora_pointers",
|
|
"sync_kt_lora_gradients",
|
|
"save_lora_experts_to_adapter",
|
|
"save_kt_moe_to_adapter",
|
|
"load_lora_experts_from_adapter",
|
|
"load_kt_moe_from_adapter",
|
|
"LoRAExpertMLP",
|
|
"LoRAExperts",
|
|
"wrap_moe_layers_with_kt_wrapper",
|
|
"build_kt_device_map",
|
|
"build_kt_device_map_simplified",
|
|
"get_kt_loading_kwargs",
|
|
"load_kt_model",
|
|
"collect_kt_sft_profile",
|
|
"format_kt_sft_profile",
|
|
"reset_kt_sft_profile",
|
|
]
|