mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-22 23:42:06 +00:00
7830de1566
This change adds support for FlashInfer. FlashInfer can be enabled using `FLASH_INFER=1` and is currently only implemented in `FlashCausalLM`. Since this functionality is currently only for testing, FlashInfer is not installed anywhere yet. The FlashInfer API is quite different from FlashAttention/vLLM in that it requires more global bookkeeping: * A wrapper class needs to be contstructed (which we just call *state*). Since this is fairly expensive (due to pinned host memory allocation), we only do this once in a FlashCausalLM instance or for each CUDA Graph size. * Each model forward call needs to be wrapped in `begin_forward` and `end_forward`. This sets up data structures that can be reused for all calls to attention for that forward call. When calling attention, we need access to the state object. To avoid passing an argument down the call chain (which would require changes to all models), we use a context variable. Each model forward call is wrapped using a context manager that does all the bookkeeping for such a call: * Set the context variable to the forward call's state. * Call `begin_forward` on the state. * Yield. * Call `end_forward` on the state. * Reset the context variable. We cannot use a single shared global variable for this, since e.g. CUDA Graphs of different sizes each have their own state.
45 lines
1.3 KiB
Python
45 lines
1.3 KiB
Python
from dataclasses import dataclass
|
|
from text_generation_server.models.globals import FLASH_DECODING, FLASH_INFER
|
|
import torch
|
|
from typing import Optional
|
|
|
|
|
|
if FLASH_DECODING or FLASH_INFER:
|
|
|
|
@dataclass
|
|
class Seqlen:
|
|
input_lengths: torch.Tensor
|
|
cu_seqlen_q: Optional[torch.Tensor]
|
|
cu_seqlen_k: Optional[torch.Tensor]
|
|
|
|
def __init__(self, input_lengths):
|
|
self.input_lengths = input_lengths
|
|
device = self.input_lengths.device
|
|
shape = self.input_lengths.shape
|
|
cu_seqlen_q = torch.arange(
|
|
shape[0] + 1,
|
|
device=device,
|
|
dtype=torch.int32,
|
|
)
|
|
cu_seqlen_k = torch.zeros(shape[-1] + 1, device=device, dtype=torch.int32)
|
|
# cuda graphs don't like this and this is necessary to clamp within mistral
|
|
# Although FA2 might not want the clamping
|
|
# cu_seqlen_k[0] = 0
|
|
torch.cumsum(self.input_lengths, -1, out=cu_seqlen_k[1:])
|
|
|
|
self.cu_seqlen_q = cu_seqlen_q
|
|
self.cu_seqlen_k = cu_seqlen_k
|
|
|
|
def clamp(self, max):
|
|
# Flash decoding doesn't need to clamp
|
|
return self
|
|
|
|
else:
|
|
|
|
@dataclass
|
|
class Seqlen:
|
|
input_lengths: torch.Tensor
|
|
|
|
def clamp(self, max):
|
|
return Seqlen(torch.clamp(self.input_lengths, max=max))
|