Do not initialize scratch space when there are no ExLlamaV2 layers (#2015)

# What does this PR do?

Do not attempt to allocate ExLlamaV2 scratch buffers when there are no
ExLlama2 layers. Avoids a crash in warmup for models that cannot use
exllama when ExLlamaV2 is installed.

## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?

## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @

@OlivierDehaene OR @Narsil

 -->
This commit is contained in:
Daniël de Kok 2024-06-05 10:45:47 +02:00 committed by yuanwu
parent 353a9669ba
commit cdd120ac02

View File

@ -1,10 +1,15 @@
# Adapted from turboderp exllama: https://github.com/turboderp/exllamav2 # Adapted from turboderp exllama: https://github.com/turboderp/exllamav2
from dataclasses import dataclass
from typing import Optional
import torch import torch
import torch.nn as nn import torch.nn as nn
from loguru import logger from loguru import logger
from text_generation_server.layers.exl2 import Exl2Weight
from text_generation_server.layers.gptq import GPTQWeight
try: try:
from exllamav2_kernels import make_q_matrix, gemm_half_q_half from exllamav2_kernels import make_q_matrix, gemm_half_q_half
except ImportError: except ImportError:
@ -15,6 +20,15 @@ except ImportError:
none_tensor = torch.empty((1, 1), device="meta") none_tensor = torch.empty((1, 1), device="meta")
@dataclass
class _ExtraTensors:
"""Additional generated quantizer tensors."""
q_group_map: Optional[torch.Tensor] = None
q_invperm: Optional[torch.Tensor] = None
q_perm: Optional[torch.Tensor] = None
def ext_gemm_half_q_half(x, q_handle, q4_width, force_cuda): def ext_gemm_half_q_half(x, q_handle, q4_width, force_cuda):
"""Matrix multiplication, returns x @ q4""" """Matrix multiplication, returns x @ q4"""
output_shape = x.shape[:-1] + (q4_width,) output_shape = x.shape[:-1] + (q4_width,)
@ -24,11 +38,7 @@ def ext_gemm_half_q_half(x, q_handle, q4_width, force_cuda):
return output.view(output_shape) return output.view(output_shape)
# Group map needed for irregular group sizes def make_group_map(q_groups: torch.Tensor, num_qrows: int):
def make_group_map(q_groups, num_qrows):
gr = q_groups.tolist() gr = q_groups.tolist()
group_map = [] group_map = []
num_groups = len(gr) // 2 num_groups = len(gr) // 2
@ -50,72 +60,72 @@ def make_group_map(q_groups, num_qrows):
# Create Q matrix # Create Q matrix
def ext_make_q_matrix(w: dict, temp_dq, key: str = None): def ext_make_q_matrix(
w: Exl2Weight | GPTQWeight,
extra: _ExtraTensors,
temp_dq,
key: Optional[str] = None,
):
""" """
Create Q matrix Create Q matrix
""" """
# EXL2 # EXL2
# won't work as the moment because the tensors are not the same. if isinstance(w, Exl2Weight):
if "q_weight" in w: extra.q_group_map = make_group_map(w.q_groups, w.q_weight.shape[0])
w["q_scale_max"] /= 256 extra.q_perm = torch.argsort(w.q_invperm).short()
w["q_perm"] = w["q_perm"].short()
w["q_invperm"] = w["q_invperm"].short()
if "q_group_map" not in w:
w["q_group_map"] = make_group_map(w["q_groups"], w["q_weight"].shape[0])
return make_q_matrix( return make_q_matrix(
w["q_weight"], w.q_weight,
w["q_perm"], extra.q_perm,
w["q_invperm"], w.q_invperm,
w["q_scale"], w.q_scale,
w["q_scale_max"], w.q_scale_max,
w["q_groups"], w.q_groups,
w["q_group_map"], extra.q_group_map,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
temp_dq, temp_dq,
) )
# GPTQ # GPTQ
elif "qweight" in w: elif isinstance(w, GPTQWeight):
if w["scales"].dtype == torch.float: if w.scales.dtype == torch.float:
w["scales"] = w["scales"].half() w.scales = w.scales.half()
# GPTQ with g_idx (act_order) # GPTQ with g_idx (act_order)
if w.get("g_idx", None) is not None and not (w["g_idx"] == 0).all().item(): if w.g_idx is not None and not (w.g_idx == 0).all().item():
w["q_perm"] = torch.empty( extra.q_perm = torch.empty(
(w["qweight"].shape[0] * 8,), (w.qweight.shape[0] * 8,),
dtype=torch.short, dtype=torch.short,
device=w["qweight"].device, device=w.qweight.device,
) )
w["q_invperm"] = torch.empty_like(w["q_perm"]) extra.q_invperm = torch.empty_like(extra.q_perm)
# make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx. # make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx.
return make_q_matrix( return make_q_matrix(
w["qweight"], w.qweight,
w["q_perm"], extra.q_perm,
w["q_invperm"], extra.q_invperm,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
w["qzeros"], w.qzeros,
w["scales"], w.scales,
w["g_idx"].cpu(), w.g_idx.cpu(),
temp_dq, temp_dq,
) )
# GPTQ without g_idx # GPTQ without g_idx
else: else:
return make_q_matrix( return make_q_matrix(
w["qweight"], w.qweight,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
none_tensor, none_tensor,
w["qzeros"], w.qzeros,
w["scales"], w.scales,
none_tensor, none_tensor,
temp_dq, temp_dq,
) )
@ -124,7 +134,6 @@ def ext_make_q_matrix(w: dict, temp_dq, key: str = None):
DEVICE = None DEVICE = None
FIXED_BYTES = 0
LAYERS = [] LAYERS = []
@ -134,8 +143,19 @@ def set_device(device):
def create_exllama_buffers(max_total_tokens: int): def create_exllama_buffers(max_total_tokens: int):
global FIXED_BYTES, LAYERS, DEVICE global LAYERS, DEVICE
temp_dq = ExLlamaV2DeviceTensors(DEVICE, FIXED_BYTES)
# No need to initialize scratch space if there are no layers
# that use ExLLamav2.
if len(LAYERS) == 0:
return
# Find the size of the scratch space.
scratch_bytes = max(
layer.scratch_space_fixed(max_input_len=max_total_tokens, max_batch_size=1)
for layer in LAYERS
)
temp_dq = ExLlamaV2DeviceTensors(DEVICE, scratch_bytes)
for layer in LAYERS: for layer in LAYERS:
layer.post_init(temp_dq) layer.post_init(temp_dq)
@ -146,49 +166,48 @@ class QuantLinear(nn.Module):
"""Linear layer implementation with per-group 4-bit quantization of the weights""" """Linear layer implementation with per-group 4-bit quantization of the weights"""
# def __init__(self, bits, group_size, infeatures, outfeatures, bias, trainable=False, **kwargs): def __init__(
def __init__(self, qweight, qzeros, scales, g_idx, bias, bits, groupsize): self,
weight: Exl2Weight | GPTQWeight,
bias: torch.Tensor,
):
super().__init__() super().__init__()
if bits != 4:
raise ValueError(
f"Exllamav2 kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization."
)
self.q_handle = None self.q_handle = None
self.q_tensors = None self.q_tensors = weight
self.bits = bits self.extra_tensors = _ExtraTensors()
self.maxq = 2**self.bits - 1
self.infeatures = qweight.shape[0] // self.bits * 32 if isinstance(weight, Exl2Weight):
self.outfeatures = qweight.shape[1] self.infeatures = weight.q_invperm.shape[0]
self.outfeatures = weight.q_weight.shape[1]
elif isinstance(weight, GPTQWeight):
if weight.bits != 4:
raise ValueError(
f"Exllamav2 kernel supports only bits=4, requested bits={weight.bits}. Something is wrong in the model initialization."
)
self.infeatures = weight.qweight.shape[0] // weight.bits * 32
self.outfeatures = weight.qweight.shape[1]
self.padding = -self.outfeatures % 32 self.padding = -self.outfeatures % 32
self.outfeatures = self.outfeatures + self.padding self.outfeatures = self.outfeatures + self.padding
self.device = qweight.device self.device = weight.device
self.qweight = qweight
self.qzeros = qzeros
self.scales = scales
self.g_idx = g_idx
self.bias = bias if bias is not None else None self.bias = bias if bias is not None else None
self.group_size = groupsize
global FIXED_BYTES, LAYERS global LAYERS
FIXED_BYTES = max(FIXED_BYTES, self.scratch_space_fixed())
LAYERS.append(self) LAYERS.append(self)
def post_init(self, temp_dq): def post_init(self, temp_dq):
assert self.qweight.device.type == "cuda" device = self.q_tensors.device
assert self.qweight.device.index is not None assert device.type == "cuda"
self.q_tensors = { assert device.index is not None
"qweight": self.qweight,
"qzeros": self.qzeros,
"scales": self.scales,
"g_idx": self.g_idx,
}
temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size()) temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size())
# We NEED to keep a pointer on Python side, otherwise the garbage collector will mess with us, # We NEED to keep a pointer on Python side, otherwise the garbage collector will mess with us,
# and `Memory access fault by GPU node-2` will EAT you. # and `Memory access fault by GPU node-2` will EAT you.
self.temp_dq = temp_dq self.temp_dq = temp_dq
self.q_handle = ext_make_q_matrix(self.q_tensors, temp_dq) self.q_handle = ext_make_q_matrix(self.q_tensors, self.extra_tensors, temp_dq)
def forward(self, x, force_cuda=False): def forward(self, x, force_cuda=False):
output = ext_gemm_half_q_half(x, self.q_handle, self.outfeatures, force_cuda) output = ext_gemm_half_q_half(x, self.q_handle, self.outfeatures, force_cuda)
@ -203,7 +222,7 @@ class QuantLinear(nn.Module):
def temp_fwd_size(self, max_input_len, max_batch_size): def temp_fwd_size(self, max_input_len, max_batch_size):
return self.outfeatures * max_input_len * max_batch_size * 4 + 128 return self.outfeatures * max_input_len * max_batch_size * 4 + 128
def scratch_space_fixed(self, max_input_len=4096, max_batch_size=16): def scratch_space_fixed(self, max_input_len, max_batch_size):
return self.temp_dq_size() + self.temp_fwd_size(max_input_len, max_batch_size) return self.temp_dq_size() + self.temp_fwd_size(max_input_len, max_batch_size)