mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-04-28 11:49:51 +00:00
57d14d22bc
* refactor: move legacy code to archive/ directory - Moved ktransformers, csrc, third_party, merge_tensors to archive/ - Moved build scripts and configurations to archive/ - Kept kt-kernel, KT-SFT, doc, and README files in root - Preserved complete git history for all moved files * refactor: restructure repository to focus on kt-kernel and KT-SFT modules * fix README * fix README * fix README * fix README * docs: add performance benchmarks to kt-kernel section Add comprehensive performance data for kt-kernel to match KT-SFT's presentation: - AMX kernel optimization: 21.3 TFLOPS (3.9× faster than PyTorch) - Prefill phase: up to 20× speedup vs baseline - Decode phase: up to 4× speedup - NUMA optimization: up to 63% throughput improvement - Multi-GPU (8×L20): 227.85 tokens/s total throughput with DeepSeek-R1 FP8 Source: https://lmsys.org/blog/2025-10-22-KTransformers/ This provides users with concrete performance metrics for both core modules, making it easier to understand the capabilities of each component. * refactor: improve kt-kernel performance data with specific hardware and models Replace generic performance descriptions with concrete benchmarks: - Specify exact hardware: 8×L20 GPU + Xeon Gold 6454S, Single/Dual-socket Xeon + AMX - Include specific models: DeepSeek-R1-0528 (FP8), DeepSeek-V3 (671B) - Show detailed metrics: total throughput, output throughput, concurrency details - Match KT-SFT presentation style for consistency This provides users with actionable performance data they can use to evaluate hardware requirements and expected performance for their use cases. * fix README * docs: clean up performance table and improve formatting * add pic for README * refactor: simplify .gitmodules and backup legacy submodules - Remove 7 legacy submodules from root .gitmodules (archive/third_party/*) - Keep only 2 active submodules for kt-kernel (llama.cpp, pybind11) - Backup complete .gitmodules to archive/.gitmodules - Add documentation in archive/README.md for researchers who need legacy submodules This reduces initial clone size by ~500MB and avoids downloading unused dependencies. * refactor: move doc/ back to root directory Keep documentation in root for easier access and maintenance. * refactor: consolidate all images to doc/assets/ - Move kt-kernel/assets/heterogeneous_computing.png to doc/assets/ - Remove KT-SFT/assets/ (images already in doc/assets/) - Update KT-SFT/README.md image references to ../doc/assets/ - Eliminates ~7.9MB image duplication - Centralizes all documentation assets in one location * fix pic path for README
83 lines
5.1 KiB
Python
83 lines
5.1 KiB
Python
from pydantic import BaseModel, Field
|
|
from typing import Optional
|
|
from ktransformers.server.config.config import Config
|
|
|
|
|
|
class ConfigArgs(BaseModel):
|
|
model_name: Optional[str] = Field(..., description="Model name")
|
|
model_dir: Optional[str] = Field(..., description="Path to model directory")
|
|
optimize_config_path: Optional[str] = Field(None, description="Path of your optimize config yml file")
|
|
gguf_path: Optional[str] = Field(None, description="Path of your gguf file")
|
|
draft_model_path: Optional[str] = Field(None, description="Path of your gguf file")
|
|
draft_gguf_path: Optional[str] = Field(None, description="Path of your gguf file")
|
|
tp: int = Field(None, description="tp size")
|
|
|
|
class Config:
|
|
protected_namespaces = ()
|
|
|
|
max_batch_size: int = Field(
|
|
None, description="Max number of batches to run at once, assuming the sequences will fit within total_context"
|
|
)
|
|
chunk_size: int = Field(
|
|
None,
|
|
description=(
|
|
"Max chunk size. Determines the size of prefill operations. Can be reduced to reduce pauses whenever a new"
|
|
" job is started, but at the expense of overall prompt ingestion speed"
|
|
),
|
|
)
|
|
max_new_tokens: int = Field(None, description="Max new tokens per completion. For this example applies to all jobs")
|
|
json_mode: bool = Field(
|
|
None, description="Use LMFE to constrain the output to JSON format. See schema and details below"
|
|
)
|
|
healing: bool = Field(None, description="Demonstrate token healing")
|
|
ban_strings: Optional[list] = Field(None, description="Ban some phrases maybe")
|
|
gpu_split: Optional[str] = Field(None, description='"auto", or VRAM allocation per GPU in GB')
|
|
length: Optional[int] = Field(None, description="Maximum sequence length")
|
|
rope_scale: Optional[float] = Field(None, description="RoPE scaling factor")
|
|
rope_alpha: Optional[float] = Field(None, description="RoPE alpha value (NTK)")
|
|
no_flash_attn: bool = Field(None, description="Disable Flash Attention")
|
|
low_mem: bool = Field(None, description="Enable VRAM optimizations, potentially trading off speed")
|
|
experts_per_token: Optional[int] = Field(
|
|
None, description="Override MoE model's default number of experts per token"
|
|
)
|
|
load_q4: bool = Field(None, description="Load weights in Q4 mode")
|
|
fast_safetensors: bool = Field(None, description="Optimized safetensors loading with direct I/O (experimental!)")
|
|
draft_model_dir: Optional[str] = Field(None, description="Path to draft model directory")
|
|
no_draft_scale: bool = Field(
|
|
None,
|
|
description="If draft model has smaller context size than model, don't apply alpha (NTK) scaling to extend it",
|
|
)
|
|
modes: bool = Field(None, description="List available modes and exit.")
|
|
mode: str = Field(None, description="Chat mode. Use llama for Llama 1/2 chat finetunes.")
|
|
username: str = Field(None, description="Username when using raw chat mode")
|
|
botname: str = Field(None, description="Bot name when using raw chat mode")
|
|
system_prompt: Optional[str] = Field(None, description="Use custom system prompt")
|
|
temperature: float = Field(None, description="Sampler temperature, default = 0.95 (1 to disable)")
|
|
smoothing_factor: float = Field(None, description="Smoothing Factor, default = 0.0 (0 to disable)")
|
|
dynamic_temperature: Optional[str] = Field(
|
|
None, description="Dynamic temperature min,max,exponent, e.g. -dyntemp 0.2,1.5,1"
|
|
)
|
|
top_k: int = Field(None, description="Sampler top-K, default = 50 (0 to disable)")
|
|
top_p: float = Field(None, description="Sampler top-P, default = 0.8 (0 to disable)")
|
|
top_a: float = Field(None, description="Sampler top-A, default = 0.0 (0 to disable)")
|
|
skew: float = Field(None, description="Skew sampling, default = 0.0 (0 to disable)")
|
|
typical: float = Field(None, description="Sampler typical threshold, default = 0.0 (0 to disable)")
|
|
repetition_penalty: float = Field(None, description="Sampler repetition penalty, default = 1.01 (1 to disable)")
|
|
frequency_penalty: float = Field(None, description="Sampler frequency penalty, default = 0.0 (0 to disable)")
|
|
presence_penalty: float = Field(None, description="Sampler presence penalty, default = 0.0 (0 to disable)")
|
|
response_chunk: int = Field(None, description="Space to reserve in context for reply, default = 250")
|
|
no_code_formatting: bool = Field(None, description="Disable code formatting/syntax highlighting")
|
|
cache_8bit: bool = Field(None, description="Use 8-bit (FP8) cache")
|
|
cache_q4: bool = Field(None, description="Use Q4 cache")
|
|
ngram_decoding: bool = Field(None, description="Use n-gram speculative decoding")
|
|
print_timings: bool = Field(None, description="Output timings after each prompt")
|
|
amnesia: bool = Field(None, description="Forget context after every response")
|
|
|
|
# for transformers
|
|
batch_size: int = Field(None, description="Batch Size")
|
|
cache_lens: int = Field(None, description="Cache lens for transformers static cache")
|
|
device: str = Field(None, description="device")
|
|
|
|
|
|
cfg = Config()
|
|
default_args = cfg
|