ef6abf3dc7
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.
95 lines
3.2 KiB
Python
95 lines
3.2 KiB
Python
from typing import Any, cast
|
|
|
|
|
|
def get_processor(
|
|
processor_name: str,
|
|
*args: Any,
|
|
trust_remote_code: bool = False,
|
|
**kwargs: Any,
|
|
):
|
|
"""Load a processor for the given model name via HuggingFace."""
|
|
# don't put this import at the top level
|
|
# it will call torch.cuda.device_count()
|
|
from transformers import AutoProcessor
|
|
from transformers.processing_utils import ProcessorMixin
|
|
|
|
try:
|
|
processor = AutoProcessor.from_pretrained(
|
|
processor_name,
|
|
*args,
|
|
trust_remote_code=trust_remote_code,
|
|
**kwargs)
|
|
except ValueError as e:
|
|
# If the error pertains to the processor class not existing or not
|
|
# currently being imported, suggest using the --trust-remote-code flag.
|
|
# Unlike AutoTokenizer, AutoProcessor does not separate such errors
|
|
if not trust_remote_code:
|
|
err_msg = (
|
|
"Failed to load the processor. If the processor is "
|
|
"a custom processor not yet available in the HuggingFace "
|
|
"transformers library, consider setting "
|
|
"`trust_remote_code=True` in LLM or using the "
|
|
"`--trust-remote-code` flag in the CLI.")
|
|
raise RuntimeError(err_msg) from e
|
|
else:
|
|
raise e
|
|
|
|
return cast(ProcessorMixin, processor)
|
|
|
|
|
|
def get_image_processor(
|
|
processor_name: str,
|
|
*args: Any,
|
|
trust_remote_code: bool = False,
|
|
**kwargs: Any,
|
|
):
|
|
"""Load an image processor for the given model name via HuggingFace."""
|
|
# don't put this import at the top level
|
|
# it will call torch.cuda.device_count()
|
|
from transformers import AutoImageProcessor
|
|
from transformers.image_processing_utils import BaseImageProcessor
|
|
|
|
try:
|
|
processor = AutoImageProcessor.from_pretrained(
|
|
processor_name,
|
|
*args,
|
|
trust_remote_code=trust_remote_code,
|
|
**kwargs)
|
|
except ValueError as e:
|
|
# If the error pertains to the processor class not existing or not
|
|
# currently being imported, suggest using the --trust-remote-code flag.
|
|
# Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
|
|
if not trust_remote_code:
|
|
err_msg = (
|
|
"Failed to load the image processor. If the image processor is "
|
|
"a custom processor not yet available in the HuggingFace "
|
|
"transformers library, consider setting "
|
|
"`trust_remote_code=True` in LLM or using the "
|
|
"`--trust-remote-code` flag in the CLI.")
|
|
raise RuntimeError(err_msg) from e
|
|
else:
|
|
raise e
|
|
|
|
return cast(BaseImageProcessor, processor)
|
|
|
|
|
|
def get_video_processor(
|
|
processor_name: str,
|
|
*args: Any,
|
|
trust_remote_code: bool = False,
|
|
**kwargs: Any,
|
|
):
|
|
"""Load a video processor for the given model name via HuggingFace."""
|
|
# don't put this import at the top level
|
|
# it will call torch.cuda.device_count()
|
|
from transformers.image_processing_utils import BaseImageProcessor
|
|
|
|
processor = get_processor(
|
|
processor_name,
|
|
*args,
|
|
trust_remote_code=trust_remote_code,
|
|
**kwargs,
|
|
)
|
|
|
|
return cast(BaseImageProcessor, processor.video_processor)
|