Files
project_6/vllm/inputs/__init__.py
dylanyunlon ef6abf3dc7 [DEPLOY] Complete submission: baseline + all optimizations
Adds ALL files needed for Dockerfile build:
  - qwen3_6_scripts/ (baseline patches + our optimizations)
  - vllm/ (full vllm package)
  - paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
  - Dockerfile + computility-run.yaml

Our optimizations vs baseline:
  1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
     Triton try/fallback, V2 heuristic, threshold 32K→64K
  2. paged_attention_v2_pytorch.py: fills NotImplementedError,
     single-bmm Phase 1 (195 launches → 3)
  3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
  4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
  5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
     max-num-batched-tokens 8192→16384

This repo can now be submitted to dev.modelhub.org.cn as-is.
2026-07-30 16:06:20 +00:00

45 lines
1.2 KiB
Python

from .data import (EncoderDecoderLLMInputs, ExplicitEncoderDecoderPrompt,
LLMInputs, PromptType, SingletonPrompt, TextPrompt,
TokensPrompt, build_explicit_enc_dec_prompt,
to_enc_dec_tuple_list, zip_enc_dec_prompts)
from .registry import InputContext, InputRegistry
INPUT_REGISTRY = InputRegistry()
"""
The global :class:`~InputRegistry` which is used by :class:`~vllm.LLMEngine`
to dispatch data processing according to the target model.
See also:
:ref:`input_processing_pipeline`
"""
__all__ = [
"TextPrompt",
"TokensPrompt",
"PromptType",
"SingletonPrompt",
"ExplicitEncoderDecoderPrompt",
"LLMInputs",
"EncoderDecoderLLMInputs",
"build_explicit_enc_dec_prompt",
"to_enc_dec_tuple_list",
"zip_enc_dec_prompts",
"INPUT_REGISTRY",
"InputContext",
"InputRegistry",
]
def __getattr__(name: str):
if name == "PromptInput":
import warnings
msg = ("PromptInput has been renamed to PromptType. "
"The original name will be removed in an upcoming version.")
warnings.warn(DeprecationWarning(msg), stacklevel=2)
return PromptType
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")