mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2026-03-16 14:27:19 +08:00
[Kernel/Quant] Remove the original marlin format and qqq (#23204)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
GitHub
parent
ebe56a0064
commit
0cdbf5e61c
@ -1,12 +0,0 @@
|
||||
# For vllm script, with -t option (tensor parallel size).
|
||||
# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
|
||||
model_name: "HandH1998/QQQ-Llama-3-8b-g128"
|
||||
tasks:
|
||||
- name: "gsm8k"
|
||||
metrics:
|
||||
- name: "exact_match,strict-match"
|
||||
value: 0.419
|
||||
- name: "exact_match,flexible-extract"
|
||||
value: 0.416
|
||||
limit: 1000
|
||||
num_fewshot: 5
|
||||
@ -3,4 +3,3 @@ Meta-Llama-3-70B-Instruct.yaml
|
||||
Mixtral-8x7B-Instruct-v0.1.yaml
|
||||
Qwen2-57B-A14-Instruct.yaml
|
||||
DeepSeek-V2-Lite-Chat.yaml
|
||||
Meta-Llama-3-8B-QQQ.yaml
|
||||
|
||||
@ -357,9 +357,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
|
||||
|
||||
set(MARLIN_SRCS
|
||||
"csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
|
||||
"csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
|
||||
"csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
|
||||
"csrc/quantization/gptq_marlin/gptq_marlin.cu"
|
||||
"csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
|
||||
"csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
|
||||
|
||||
@ -253,28 +253,7 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
|
||||
else:
|
||||
assert bt.a.dtype == torch.int8
|
||||
assert bt.wtype == scalar_types.uint4b8
|
||||
|
||||
if bt.w_ch_s is not None:
|
||||
s_ch = bt.w_ch_s.to(torch.float32)
|
||||
else:
|
||||
s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device)
|
||||
|
||||
if bt.w_tok_s is not None:
|
||||
s_tok = bt.w_tok_s.to(torch.float32)
|
||||
else:
|
||||
s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device)
|
||||
|
||||
fn = lambda: ops.marlin_qqq_gemm(
|
||||
a=bt.a,
|
||||
b_q_weight=w_q,
|
||||
s_group=w_s,
|
||||
s_tok=s_tok,
|
||||
s_ch=s_ch,
|
||||
workspace=workspace.scratch,
|
||||
size_m=bt.a.shape[0],
|
||||
size_n=bt.w_ref.shape[1],
|
||||
size_k=bt.w_ref.shape[0],
|
||||
)
|
||||
raise NotImplementedError("QQQ is not supported anymore")
|
||||
|
||||
return fn
|
||||
|
||||
|
||||
@ -571,78 +571,79 @@ def generate():
|
||||
itertools.repeat(default_heuristic))
|
||||
]
|
||||
|
||||
# Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
|
||||
# TODO (LucasWilkinson): Further tuning required
|
||||
qqq_tile_heuristic_config = {
|
||||
#### M = 257+
|
||||
# ((128, 256), (2, 1, 1)) Broken for QQQ types
|
||||
# TODO (LucasWilkinson): Investigate further
|
||||
# "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
|
||||
# "M > 256": ((128, 256), (2, 1, 1)),
|
||||
"M > 256": ((128, 128), (2, 1, 1)),
|
||||
#### M = 129-256
|
||||
"M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
|
||||
"M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
|
||||
# ((128, 256), (2, 1, 1)) Broken for QQQ types
|
||||
# TODO (LucasWilkinson): Investigate further
|
||||
# "M > 128": ((128, 256), (2, 1, 1)),
|
||||
"M > 128": ((128, 128), (2, 1, 1)),
|
||||
#### M = 65-128
|
||||
"M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
|
||||
"M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
|
||||
"M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
|
||||
"M > 64": ((128, 128), (2, 1, 1)),
|
||||
#### M = 33-64
|
||||
"M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
|
||||
# Broken for QQQ types
|
||||
# TODO (LucasWilkinson): Investigate further
|
||||
#"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
|
||||
"M > 32": ((128, 64), (2, 1, 1)),
|
||||
#### M = 17-32
|
||||
"M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
|
||||
"M > 16": ((256, 32), (2, 1, 1)),
|
||||
#### M = 1-16
|
||||
"N >= 26624": ((256, 16), (1, 1, 1)),
|
||||
None: ((128, 16), (1, 1, 1)),
|
||||
}
|
||||
# TODO: Support W4A8 when ready
|
||||
# # Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
|
||||
# # TODO (LucasWilkinson): Further tuning required
|
||||
# qqq_tile_heuristic_config = {
|
||||
# #### M = 257+
|
||||
# # ((128, 256), (2, 1, 1)) Broken for QQQ types
|
||||
# # TODO (LucasWilkinson): Investigate further
|
||||
# # "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
|
||||
# # "M > 256": ((128, 256), (2, 1, 1)),
|
||||
# "M > 256": ((128, 128), (2, 1, 1)),
|
||||
# #### M = 129-256
|
||||
# "M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
|
||||
# "M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
|
||||
# # ((128, 256), (2, 1, 1)) Broken for QQQ types
|
||||
# # TODO (LucasWilkinson): Investigate further
|
||||
# # "M > 128": ((128, 256), (2, 1, 1)),
|
||||
# "M > 128": ((128, 128), (2, 1, 1)),
|
||||
# #### M = 65-128
|
||||
# "M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
|
||||
# "M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
|
||||
# "M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
|
||||
# "M > 64": ((128, 128), (2, 1, 1)),
|
||||
# #### M = 33-64
|
||||
# "M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
|
||||
# # Broken for QQQ types
|
||||
# # TODO (LucasWilkinson): Investigate further
|
||||
# #"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
|
||||
# "M > 32": ((128, 64), (2, 1, 1)),
|
||||
# #### M = 17-32
|
||||
# "M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
|
||||
# "M > 16": ((256, 32), (2, 1, 1)),
|
||||
# #### M = 1-16
|
||||
# "N >= 26624": ((256, 16), (1, 1, 1)),
|
||||
# None: ((128, 16), (1, 1, 1)),
|
||||
# }
|
||||
|
||||
# For now we use the same heuristic for all types
|
||||
# Heuristic is currently tuned for H100s
|
||||
qqq_heuristic = [
|
||||
(cond, ScheduleConfig(*tile_config,
|
||||
**sch_common_params)) # type: ignore
|
||||
for cond, tile_config in qqq_tile_heuristic_config.items()
|
||||
]
|
||||
# # For now we use the same heuristic for all types
|
||||
# # Heuristic is currently tuned for H100s
|
||||
# qqq_heuristic = [
|
||||
# (cond, ScheduleConfig(*tile_config,
|
||||
# **sch_common_params)) # type: ignore
|
||||
# for cond, tile_config in qqq_tile_heuristic_config.items()
|
||||
# ]
|
||||
|
||||
QQQ_kernel_types = [
|
||||
*(TypeConfig(
|
||||
a=DataType.s8,
|
||||
b=VLLMDataType.u4b8,
|
||||
b_group_scale=b_group_scale,
|
||||
b_group_zeropoint=DataType.void,
|
||||
b_channel_scale=DataType.f32,
|
||||
a_token_scale=DataType.f32,
|
||||
out=DataType.f16,
|
||||
accumulator=DataType.s32,
|
||||
) for b_group_scale in (DataType.f16, DataType.void)),
|
||||
*(TypeConfig(
|
||||
a=DataType.e4m3,
|
||||
b=VLLMDataType.u4b8,
|
||||
b_group_scale=b_group_scale,
|
||||
b_group_zeropoint=DataType.void,
|
||||
b_channel_scale=DataType.f32,
|
||||
a_token_scale=DataType.f32,
|
||||
out=DataType.f16,
|
||||
accumulator=DataType.f32,
|
||||
) for b_group_scale in (DataType.f16, DataType.void)),
|
||||
]
|
||||
# QQQ_kernel_types = [
|
||||
# *(TypeConfig(
|
||||
# a=DataType.s8,
|
||||
# b=VLLMDataType.u4b8,
|
||||
# b_group_scale=b_group_scale,
|
||||
# b_group_zeropoint=DataType.void,
|
||||
# b_channel_scale=DataType.f32,
|
||||
# a_token_scale=DataType.f32,
|
||||
# out=DataType.f16,
|
||||
# accumulator=DataType.s32,
|
||||
# ) for b_group_scale in (DataType.f16, DataType.void)),
|
||||
# *(TypeConfig(
|
||||
# a=DataType.e4m3,
|
||||
# b=VLLMDataType.u4b8,
|
||||
# b_group_scale=b_group_scale,
|
||||
# b_group_zeropoint=DataType.void,
|
||||
# b_channel_scale=DataType.f32,
|
||||
# a_token_scale=DataType.f32,
|
||||
# out=DataType.f16,
|
||||
# accumulator=DataType.f32,
|
||||
# ) for b_group_scale in (DataType.f16, DataType.void)),
|
||||
# ]
|
||||
|
||||
impl_configs += [
|
||||
ImplConfig(x[0], x[1], x[2])
|
||||
for x in zip(QQQ_kernel_types,
|
||||
itertools.repeat(get_unique_schedules(qqq_heuristic)),
|
||||
itertools.repeat(qqq_heuristic))
|
||||
]
|
||||
# impl_configs += [
|
||||
# ImplConfig(x[0], x[1], x[2])
|
||||
# for x in zip(QQQ_kernel_types,
|
||||
# itertools.repeat(get_unique_schedules(qqq_heuristic)),
|
||||
# itertools.repeat(qqq_heuristic))
|
||||
# ]
|
||||
|
||||
output_dir = os.path.join(SCRIPT_DIR, "generated")
|
||||
|
||||
|
||||
@ -1,209 +0,0 @@
|
||||
Contains code from https://github.com/IST-DASLab/marlin
|
||||
|
||||
Apache License
|
||||
Version 2.0, January 2004
|
||||
http://www.apache.org/licenses/
|
||||
|
||||
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
||||
|
||||
1. Definitions.
|
||||
|
||||
"License" shall mean the terms and conditions for use, reproduction,
|
||||
and distribution as defined by Sections 1 through 9 of this document.
|
||||
|
||||
"Licensor" shall mean the copyright owner or entity authorized by
|
||||
the copyright owner that is granting the License.
|
||||
|
||||
"Legal Entity" shall mean the union of the acting entity and all
|
||||
other entities that control, are controlled by, or are under common
|
||||
control with that entity. For the purposes of this definition,
|
||||
"control" means (i) the power, direct or indirect, to cause the
|
||||
direction or management of such entity, whether by contract or
|
||||
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
||||
outstanding shares, or (iii) beneficial ownership of such entity.
|
||||
|
||||
"You" (or "Your") shall mean an individual or Legal Entity
|
||||
exercising permissions granted by this License.
|
||||
|
||||
"Source" form shall mean the preferred form for making modifications,
|
||||
including but not limited to software source code, documentation
|
||||
source, and configuration files.
|
||||
|
||||
"Object" form shall mean any form resulting from mechanical
|
||||
transformation or translation of a Source form, including but
|
||||
not limited to compiled object code, generated documentation,
|
||||
and conversions to other media types.
|
||||
|
||||
"Work" shall mean the work of authorship, whether in Source or
|
||||
Object form, made available under the License, as indicated by a
|
||||
copyright notice that is included in or attached to the work
|
||||
(an example is provided in the Appendix below).
|
||||
|
||||
"Derivative Works" shall mean any work, whether in Source or Object
|
||||
form, that is based on (or derived from) the Work and for which the
|
||||
editorial revisions, annotations, elaborations, or other modifications
|
||||
represent, as a whole, an original work of authorship. For the purposes
|
||||
of this License, Derivative Works shall not include works that remain
|
||||
separable from, or merely link (or bind by name) to the interfaces of,
|
||||
the Work and Derivative Works thereof.
|
||||
|
||||
"Contribution" shall mean any work of authorship, including
|
||||
the original version of the Work and any modifications or additions
|
||||
to that Work or Derivative Works thereof, that is intentionally
|
||||
submitted to Licensor for inclusion in the Work by the copyright owner
|
||||
or by an individual or Legal Entity authorized to submit on behalf of
|
||||
the copyright owner. For the purposes of this definition, "submitted"
|
||||
means any form of electronic, verbal, or written communication sent
|
||||
to the Licensor or its representatives, including but not limited to
|
||||
communication on electronic mailing lists, source code control systems,
|
||||
and issue tracking systems that are managed by, or on behalf of, the
|
||||
Licensor for the purpose of discussing and improving the Work, but
|
||||
excluding communication that is conspicuously marked or otherwise
|
||||
designated in writing by the copyright owner as "Not a Contribution."
|
||||
|
||||
"Contributor" shall mean Licensor and any individual or Legal Entity
|
||||
on behalf of whom a Contribution has been received by Licensor and
|
||||
subsequently incorporated within the Work.
|
||||
|
||||
2. Grant of Copyright License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
copyright license to reproduce, prepare Derivative Works of,
|
||||
publicly display, publicly perform, sublicense, and distribute the
|
||||
Work and such Derivative Works in Source or Object form.
|
||||
|
||||
3. Grant of Patent License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
(except as stated in this section) patent license to make, have made,
|
||||
use, offer to sell, sell, import, and otherwise transfer the Work,
|
||||
where such license applies only to those patent claims licensable
|
||||
by such Contributor that are necessarily infringed by their
|
||||
Contribution(s) alone or by combination of their Contribution(s)
|
||||
with the Work to which such Contribution(s) was submitted. If You
|
||||
institute patent litigation against any entity (including a
|
||||
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
||||
or a Contribution incorporated within the Work constitutes direct
|
||||
or contributory patent infringement, then any patent licenses
|
||||
granted to You under this License for that Work shall terminate
|
||||
as of the date such litigation is filed.
|
||||
|
||||
4. Redistribution. You may reproduce and distribute copies of the
|
||||
Work or Derivative Works thereof in any medium, with or without
|
||||
modifications, and in Source or Object form, provided that You
|
||||
meet the following conditions:
|
||||
|
||||
(a) You must give any other recipients of the Work or
|
||||
Derivative Works a copy of this License; and
|
||||
|
||||
(b) You must cause any modified files to carry prominent notices
|
||||
stating that You changed the files; and
|
||||
|
||||
(c) You must retain, in the Source form of any Derivative Works
|
||||
that You distribute, all copyright, patent, trademark, and
|
||||
attribution notices from the Source form of the Work,
|
||||
excluding those notices that do not pertain to any part of
|
||||
the Derivative Works; and
|
||||
|
||||
(d) If the Work includes a "NOTICE" text file as part of its
|
||||
distribution, then any Derivative Works that You distribute must
|
||||
include a readable copy of the attribution notices contained
|
||||
within such NOTICE file, excluding those notices that do not
|
||||
pertain to any part of the Derivative Works, in at least one
|
||||
of the following places: within a NOTICE text file distributed
|
||||
as part of the Derivative Works; within the Source form or
|
||||
documentation, if provided along with the Derivative Works; or,
|
||||
within a display generated by the Derivative Works, if and
|
||||
wherever such third-party notices normally appear. The contents
|
||||
of the NOTICE file are for informational purposes only and
|
||||
do not modify the License. You may add Your own attribution
|
||||
notices within Derivative Works that You distribute, alongside
|
||||
or as an addendum to the NOTICE text from the Work, provided
|
||||
that such additional attribution notices cannot be construed
|
||||
as modifying the License.
|
||||
|
||||
You may add Your own copyright statement to Your modifications and
|
||||
may provide additional or different license terms and conditions
|
||||
for use, reproduction, or distribution of Your modifications, or
|
||||
for any such Derivative Works as a whole, provided Your use,
|
||||
reproduction, and distribution of the Work otherwise complies with
|
||||
the conditions stated in this License.
|
||||
|
||||
5. Submission of Contributions. Unless You explicitly state otherwise,
|
||||
any Contribution intentionally submitted for inclusion in the Work
|
||||
by You to the Licensor shall be under the terms and conditions of
|
||||
this License, without any additional terms or conditions.
|
||||
Notwithstanding the above, nothing herein shall supersede or modify
|
||||
the terms of any separate license agreement you may have executed
|
||||
with Licensor regarding such Contributions.
|
||||
|
||||
6. Trademarks. This License does not grant permission to use the trade
|
||||
names, trademarks, service marks, or product names of the Licensor,
|
||||
except as required for reasonable and customary use in describing the
|
||||
origin of the Work and reproducing the content of the NOTICE file.
|
||||
|
||||
7. Disclaimer of Warranty. Unless required by applicable law or
|
||||
agreed to in writing, Licensor provides the Work (and each
|
||||
Contributor provides its Contributions) on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
||||
implied, including, without limitation, any warranties or conditions
|
||||
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
||||
PARTICULAR PURPOSE. You are solely responsible for determining the
|
||||
appropriateness of using or redistributing the Work and assume any
|
||||
risks associated with Your exercise of permissions under this License.
|
||||
|
||||
8. Limitation of Liability. In no event and under no legal theory,
|
||||
whether in tort (including negligence), contract, or otherwise,
|
||||
unless required by applicable law (such as deliberate and grossly
|
||||
negligent acts) or agreed to in writing, shall any Contributor be
|
||||
liable to You for damages, including any direct, indirect, special,
|
||||
incidental, or consequential damages of any character arising as a
|
||||
result of this License or out of the use or inability to use the
|
||||
Work (including but not limited to damages for loss of goodwill,
|
||||
work stoppage, computer failure or malfunction, or any and all
|
||||
other commercial damages or losses), even if such Contributor
|
||||
has been advised of the possibility of such damages.
|
||||
|
||||
9. Accepting Warranty or Additional Liability. While redistributing
|
||||
the Work or Derivative Works thereof, You may choose to offer,
|
||||
and charge a fee for, acceptance of support, warranty, indemnity,
|
||||
or other liability obligations and/or rights consistent with this
|
||||
License. However, in accepting such obligations, You may act only
|
||||
on Your own behalf and on Your sole responsibility, not on behalf
|
||||
of any other Contributor, and only if You agree to indemnify,
|
||||
defend, and hold each Contributor harmless for any liability
|
||||
incurred by, or claims asserted against, such Contributor by reason
|
||||
of your accepting any such warranty or additional liability.
|
||||
|
||||
END OF TERMS AND CONDITIONS
|
||||
|
||||
APPENDIX: How to apply the Apache License to your work.
|
||||
|
||||
To apply the Apache License to your work, attach the following
|
||||
boilerplate notice, with the fields enclosed by brackets "{}"
|
||||
replaced with your own identifying information. (Don't include
|
||||
the brackets!) The text should be enclosed in the appropriate
|
||||
comment syntax for the file format. We also recommend that a
|
||||
file or class name and description of purpose be included on the
|
||||
same "printed page" as the copyright notice for easier
|
||||
identification within third-party archives.
|
||||
|
||||
Copyright {yyyy} {name of copyright owner}
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
||||
------------------------------------------------------------------------------------
|
||||
|
||||
This product bundles various third-party components under other open source licenses.
|
||||
This section summarizes those components and their licenses. See licenses/
|
||||
for text of these licenses.
|
||||
@ -1,32 +0,0 @@
|
||||
/*
|
||||
* Modified by HandH1998
|
||||
* Modified by Neural Magic
|
||||
* Copyright (C) Marlin.2024 Elias Frantar
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
|
||||
|
||||
// Instances of `Vec` are used to organize groups of >>registers<<, as needed
|
||||
// for instance as inputs to tensor core operations. Consequently, all
|
||||
// corresponding index accesses must be compile-time constants, which is why we
|
||||
// extensively use `#pragma unroll` throughout the kernel code to guarantee
|
||||
// this.
|
||||
template <typename T, int n>
|
||||
struct Vec {
|
||||
T elems[n];
|
||||
__device__ T& operator[](int i) { return elems[i]; }
|
||||
};
|
||||
@ -1,89 +0,0 @@
|
||||
/*
|
||||
* Modified by HandH1998
|
||||
* Modified by Neural Magic
|
||||
* Copyright (C) Marlin.2024 Elias Frantar
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
// Predicated asynchronous global->shared copy; used for inputs A where we apply
|
||||
// predication to handle batchsizes that are not multiples of 16.
|
||||
__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
|
||||
bool pred = true) {
|
||||
const int BYTES = 16;
|
||||
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
|
||||
asm volatile(
|
||||
"{\n"
|
||||
" .reg .pred p;\n"
|
||||
" setp.ne.b32 p, %0, 0;\n"
|
||||
" @p cp.async.cg.shared.global [%1], [%2], %3;\n"
|
||||
"}\n" ::"r"((int)pred),
|
||||
"r"(smem), "l"(glob_ptr), "n"(BYTES));
|
||||
}
|
||||
|
||||
// Asynchronous global->shared copy
|
||||
__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
|
||||
const int BYTES = 16;
|
||||
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
|
||||
asm volatile(
|
||||
"{\n"
|
||||
" cp.async.cg.shared.global [%0], [%1], %2;\n"
|
||||
"}\n" ::"r"(smem),
|
||||
"l"(glob_ptr), "n"(BYTES));
|
||||
}
|
||||
|
||||
// Async copy fence.
|
||||
__device__ inline void cp_async_fence() {
|
||||
asm volatile("cp.async.commit_group;\n" ::);
|
||||
}
|
||||
|
||||
// Wait until at most `n` async copy stages are still pending.
|
||||
template <int n>
|
||||
__device__ inline void cp_async_wait() {
|
||||
asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
|
||||
}
|
||||
|
||||
// Wait until barrier reaches `count`, then lock for current threadblock.
|
||||
__device__ inline void barrier_acquire(int* lock, int count) {
|
||||
if (threadIdx.x == 0) {
|
||||
int state = -1;
|
||||
do
|
||||
// Guarantee that subsequent writes by this threadblock will be visible
|
||||
// globally.
|
||||
asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n"
|
||||
: "=r"(state)
|
||||
: "l"(lock));
|
||||
while (state != count);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Release barrier and increment visitation count.
|
||||
__device__ inline void barrier_release(int* lock, bool reset = false) {
|
||||
__syncthreads();
|
||||
if (threadIdx.x == 0) {
|
||||
if (reset) {
|
||||
lock[0] = 0;
|
||||
return;
|
||||
}
|
||||
int val = 1;
|
||||
// Make sure that all writes since acquiring this barrier are visible
|
||||
// globally, while releasing the barrier.
|
||||
asm volatile("fence.acq_rel.gpu;\n");
|
||||
asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n"
|
||||
:
|
||||
: "l"(lock), "r"(val));
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -241,14 +241,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
|
||||
// custom types:
|
||||
// https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
|
||||
|
||||
// Marlin (Dense) Optimized Quantized GEMM for GPTQ.
|
||||
ops.def(
|
||||
"marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
|
||||
"Tensor! workspace, SymInt size_m, SymInt size_n, SymInt size_k) -> "
|
||||
"Tensor",
|
||||
{stride_tag});
|
||||
// conditionally compiled so impl in source file
|
||||
|
||||
// Marlin_24 (Sparse) Optimized Quantized GEMM for GPTQ.
|
||||
ops.def(
|
||||
"gptq_marlin_24_gemm(Tensor a, Tensor b_q_weight, Tensor b_meta, "
|
||||
@ -353,15 +345,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
|
||||
ops.def("ggml_moe_get_block_size", &ggml_moe_get_block_size);
|
||||
|
||||
#ifndef USE_ROCM
|
||||
// marlin_qqq_gemm for QQQ.
|
||||
ops.def(
|
||||
"marlin_qqq_gemm(Tensor a, Tensor b_q_weight, "
|
||||
"Tensor s_tok, Tensor s_ch, Tensor s_group, "
|
||||
"Tensor! workspace, SymInt size_m, SymInt size_n, "
|
||||
"SymInt size_k) -> Tensor",
|
||||
{stride_tag});
|
||||
// conditionally compiled so impl registration is in source file
|
||||
|
||||
// CUTLASS nvfp4 block scaled GEMM
|
||||
ops.def(
|
||||
"cutlass_scaled_fp4_mm(Tensor! out, Tensor a, Tensor b,"
|
||||
|
||||
@ -53,12 +53,6 @@ def models_list(*, all: bool = True, keywords: Optional[list[str]] = None):
|
||||
"quantization": "gptq_marlin_24"
|
||||
}))
|
||||
|
||||
if is_quant_method_supported("marlin"):
|
||||
TEST_MODELS.append(
|
||||
("robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin", {
|
||||
"quantization": "marlin"
|
||||
}))
|
||||
|
||||
if not current_platform.is_rocm() and is_quant_method_supported("awq"):
|
||||
TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", {
|
||||
"quantization": "AWQ"
|
||||
|
||||
@ -95,23 +95,23 @@ TEST_TYPES = [
|
||||
token_scale_type=None)
|
||||
for w_type in [scalar_types.uint4, scalar_types.uint8]
|
||||
for a_type in [torch.float16, torch.bfloat16]),
|
||||
# QQQ style
|
||||
*(TypeConfig(act_type=torch.int8,
|
||||
weight_type=scalar_types.uint4b8,
|
||||
output_type=torch.float16,
|
||||
group_scale_type=group_scale_type,
|
||||
group_zero_type=None,
|
||||
channel_scale_type=torch.float,
|
||||
token_scale_type=torch.float)
|
||||
for group_scale_type in [None, torch.float16]),
|
||||
*(TypeConfig(act_type=torch.float8_e4m3fn,
|
||||
weight_type=scalar_types.uint4b8,
|
||||
output_type=torch.float16,
|
||||
group_scale_type=group_scale_type,
|
||||
group_zero_type=None,
|
||||
channel_scale_type=torch.float,
|
||||
token_scale_type=torch.float)
|
||||
for group_scale_type in [None, torch.float16]),
|
||||
# # QQQ style
|
||||
# *(TypeConfig(act_type=torch.int8,
|
||||
# weight_type=scalar_types.uint4b8,
|
||||
# output_type=torch.float16,
|
||||
# group_scale_type=group_scale_type,
|
||||
# group_zero_type=None,
|
||||
# channel_scale_type=torch.float,
|
||||
# token_scale_type=torch.float)
|
||||
# for group_scale_type in [None, torch.float16]),
|
||||
# *(TypeConfig(act_type=torch.float8_e4m3fn,
|
||||
# weight_type=scalar_types.uint4b8,
|
||||
# output_type=torch.float16,
|
||||
# group_scale_type=group_scale_type,
|
||||
# group_zero_type=None,
|
||||
# channel_scale_type=torch.float,
|
||||
# token_scale_type=torch.float)
|
||||
# for group_scale_type in [None, torch.float16]),
|
||||
]
|
||||
|
||||
# TODO: in future PR refactor this and `is_quant_method_supported` in the kernel
|
||||
|
||||
@ -13,11 +13,7 @@ from vllm import _custom_ops as ops
|
||||
from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
|
||||
GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
|
||||
GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
|
||||
from vllm.model_executor.layers.quantization.qqq import (
|
||||
MARLIN_QQQ_MAX_PARALLEL, MARLIN_QQQ_MIN_THREAD_N,
|
||||
MARLIN_QQQ_SUPPORTED_GROUP_SIZES, MARLIN_QQQ_SUPPORTED_NUM_BITS)
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
|
||||
GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
|
||||
MARLIN_SUPPORTED_GROUP_SIZES, marlin_make_empty_g_idx,
|
||||
marlin_make_workspace_new, marlin_permute_bias, marlin_permute_scales,
|
||||
query_marlin_supported_quant_types)
|
||||
@ -31,8 +27,6 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
|
||||
marlin_weights)
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
|
||||
marlin_24_quantize)
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils_test_qqq import ( # noqa: E501
|
||||
marlin_qqq_quantize)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
awq_pack, gptq_pack, gptq_quantize_weights, quantize_weights, sort_weights)
|
||||
from vllm.scalar_type import scalar_types
|
||||
@ -449,68 +443,6 @@ def test_hqq_marlin_gemm(
|
||||
assert max_diff < 0.04
|
||||
|
||||
|
||||
@pytest.mark.skipif(not is_quant_method_supported("qqq"),
|
||||
reason="Marlin is not supported on this GPU type.")
|
||||
@pytest.mark.parametrize("k_chunk", MARLIN_K_CHUNKS)
|
||||
@pytest.mark.parametrize("n_chunk", MARLIN_N_CHUNKS)
|
||||
@pytest.mark.parametrize("num_bits", MARLIN_QQQ_SUPPORTED_NUM_BITS)
|
||||
@pytest.mark.parametrize("group_size", MARLIN_QQQ_SUPPORTED_GROUP_SIZES)
|
||||
@pytest.mark.parametrize("mnk_factors", MNK_FACTORS)
|
||||
def test_marlin_qqq_gemm(
|
||||
k_chunk,
|
||||
n_chunk,
|
||||
num_bits,
|
||||
group_size,
|
||||
mnk_factors,
|
||||
):
|
||||
int8_traits = torch.iinfo(torch.int8)
|
||||
m_factor, n_factor, k_factor = mnk_factors
|
||||
|
||||
size_m = m_factor
|
||||
size_k = k_chunk * k_factor
|
||||
size_n = n_chunk * n_factor
|
||||
|
||||
a_input = rand_data((size_m, size_k))
|
||||
b_weight = rand_data((size_k, size_n))
|
||||
|
||||
# Quantize activations
|
||||
s_a = a_input.abs().max(dim=-1, keepdim=True)[0].div(int8_traits.max).to(
|
||||
torch.float)
|
||||
q_a = (a_input / s_a).round().clamp(int8_traits.min,
|
||||
int8_traits.max).to(torch.int8)
|
||||
|
||||
# Quantize weights
|
||||
w_ref, marlin_qqq_q_w, marlin_qqq_s_group, marlin_qqq_s_channel = \
|
||||
marlin_qqq_quantize(b_weight, num_bits, group_size)
|
||||
|
||||
workspace = MarlinWorkspace(size_n, MARLIN_QQQ_MIN_THREAD_N,
|
||||
MARLIN_QQQ_MAX_PARALLEL)
|
||||
|
||||
opcheck(torch.ops._C.marlin_qqq_gemm,
|
||||
(q_a, marlin_qqq_q_w, s_a, marlin_qqq_s_channel,
|
||||
marlin_qqq_s_group, workspace.scratch, a_input.shape[0],
|
||||
b_weight.shape[1], a_input.shape[1]))
|
||||
|
||||
output = ops.marlin_qqq_gemm(
|
||||
q_a,
|
||||
marlin_qqq_q_w,
|
||||
s_a,
|
||||
marlin_qqq_s_channel,
|
||||
marlin_qqq_s_group,
|
||||
workspace.scratch,
|
||||
a_input.shape[0],
|
||||
b_weight.shape[1],
|
||||
a_input.shape[1],
|
||||
)
|
||||
output_ref = torch.matmul(q_a.half() * s_a.half(), w_ref)
|
||||
|
||||
torch.cuda.synchronize()
|
||||
|
||||
max_diff = compute_max_diff(output, output_ref)
|
||||
|
||||
assert max_diff < 0.04
|
||||
|
||||
|
||||
def test_marlin_gemm_subset_input():
|
||||
quant_type = scalar_types.uint4b8
|
||||
group_size = 128
|
||||
@ -602,18 +534,3 @@ def test_marlin_gemm_with_bias(size_m):
|
||||
max_diff = compute_max_diff(output, output_ref)
|
||||
|
||||
assert max_diff < 0.04
|
||||
|
||||
|
||||
def test_marlin_gemm_opcheck():
|
||||
size_m = 2048
|
||||
size_n = 4096
|
||||
size_k = 4096
|
||||
a = torch.rand((size_m, size_n), device='cuda', dtype=torch.float16)
|
||||
w = torch.randint(-5, 5, (256, 8192), device='cuda', dtype=torch.int32)
|
||||
s = torch.full((32, size_k), 0.125, device='cuda', dtype=torch.float16)
|
||||
wk = MarlinWorkspace(size_n, GPTQ_MARLIN_MIN_THREAD_N,
|
||||
GPTQ_MARLIN_MAX_PARALLEL).scratch
|
||||
x = torch.ops._C.marlin_gemm(a, w, s, wk, size_m, size_n, size_k)
|
||||
y = torch.ops._C.marlin_gemm(a, w, s, wk, size_m, size_n, size_k)
|
||||
torch.testing.assert_close(x, y)
|
||||
opcheck(torch.ops._C.marlin_gemm, (a, w, s, wk, size_m, size_n, size_k))
|
||||
|
||||
@ -22,22 +22,12 @@ class ModelPair:
|
||||
MODEL_ARG_EXPTYPES = [
|
||||
# AUTOGPTQ
|
||||
# compat: autogptq <=0.7.1 is_marlin_format: bool
|
||||
# Model Serialized in Marlin Format should always use Marlin kernel.
|
||||
("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", None, "marlin"),
|
||||
("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "marlin", "marlin"),
|
||||
("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "gptq", "marlin"),
|
||||
("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "awq", "ERROR"),
|
||||
# Model Serialized in Exllama Format.
|
||||
("TheBloke/Llama-2-7B-Chat-GPTQ", None, "gptq_marlin"),
|
||||
("TheBloke/Llama-2-7B-Chat-GPTQ", "marlin", "gptq_marlin"),
|
||||
("TheBloke/Llama-2-7B-Chat-GPTQ", "gptq", "gptq"),
|
||||
("TheBloke/Llama-2-7B-Chat-GPTQ", "awq", "ERROR"),
|
||||
# compat: autogptq >=0.8.0 use checkpoint_format: str
|
||||
# Model Serialized in Marlin Format should always use Marlin kernel.
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", None, "marlin"),
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "marlin", "marlin"),
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "gptq", "marlin"),
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "awq", "ERROR"),
|
||||
# Model Serialized in Exllama Format.
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", None, "gptq_marlin"),
|
||||
("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "marlin", "gptq_marlin"),
|
||||
|
||||
@ -11,7 +11,6 @@ import torch
|
||||
from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
|
||||
from vllm.model_executor.layers.quantization.gptq_marlin import (
|
||||
GPTQMarlinLinearMethod)
|
||||
from vllm.model_executor.layers.quantization.marlin import MarlinLinearMethod
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
UnquantizedEmbeddingMethod)
|
||||
|
||||
@ -19,9 +18,7 @@ PROMPT = "On the surface of Mars, we found"
|
||||
|
||||
MODELS_QUANT = [
|
||||
("ModelCloud/Qwen1.5-1.8B-Chat-GPTQ-4bits-dynamic-cfg-with-lm_head", True),
|
||||
("ModelCloud/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit-10-25-2024", False),
|
||||
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", False),
|
||||
("neuralmagic/Meta-Llama-3-8B-Instruct-FP8", False)
|
||||
]
|
||||
|
||||
|
||||
@ -41,8 +38,7 @@ def test_lm_head(
|
||||
lm_head_layer = model.lm_head
|
||||
if lm_head_quantized:
|
||||
assert isinstance(lm_head_layer.quant_method,
|
||||
(GPTQLinearMethod, GPTQMarlinLinearMethod,
|
||||
MarlinLinearMethod))
|
||||
(GPTQLinearMethod, GPTQMarlinLinearMethod))
|
||||
else:
|
||||
assert isinstance(lm_head_layer.quant_method,
|
||||
UnquantizedEmbeddingMethod)
|
||||
|
||||
@ -26,9 +26,5 @@ compressed-tensors, nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-W8A8-testing
|
||||
awq, casperhansen/mixtral-instruct-awq, main
|
||||
awq_marlin, casperhansen/mixtral-instruct-awq, main
|
||||
fp8, neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV, main
|
||||
marlin, nm-testing/zephyr-beta-7b-marlin-g128, main
|
||||
marlin, robertgshaw2/zephyr-7b-beta-channelwise-marlin, main
|
||||
qqq, HandH1998/QQQ-Llama-3-8b-g128, main
|
||||
qqq, HandH1998/QQQ-Llama-3-8b, main
|
||||
hqq, nm-testing/Llama-3.2-1B-Instruct-HQQ, main
|
||||
None, mgleize/fairseq2-dummy-Llama-3.2-1B, main
|
||||
@ -387,14 +387,6 @@ def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
|
||||
torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
|
||||
|
||||
|
||||
# marlin
|
||||
def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
||||
b_scales: torch.Tensor, workspace: torch.Tensor, size_m: int,
|
||||
size_n: int, size_k: int) -> torch.Tensor:
|
||||
return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
|
||||
size_n, size_k)
|
||||
|
||||
|
||||
# marlin_24
|
||||
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
||||
b_meta: torch.Tensor, b_scales: torch.Tensor,
|
||||
@ -437,25 +429,6 @@ if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
|
||||
is_zp_float: bool = False) -> torch.Tensor:
|
||||
return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
|
||||
|
||||
@register_fake("_C::marlin_qqq_gemm")
|
||||
def _marlin_qqq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
|
||||
s_tok: torch.Tensor, s_ch: torch.Tensor,
|
||||
s_group: torch.Tensor, workspace: torch.Tensor,
|
||||
size_m: torch.SymInt, size_n: torch.SymInt,
|
||||
size_k: torch.SymInt) -> torch.Tensor:
|
||||
return torch.empty((size_m, size_n),
|
||||
dtype=torch.float16,
|
||||
device=a.device)
|
||||
|
||||
@register_fake("_C::marlin_gemm")
|
||||
def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
|
||||
b_scales: torch.Tensor, workspace: torch.Tensor,
|
||||
size_m: torch.SymInt, size_n: torch.SymInt,
|
||||
size_k: torch.SymInt) -> torch.Tensor:
|
||||
return torch.empty((size_m, size_n),
|
||||
dtype=torch.float16,
|
||||
device=a.device)
|
||||
|
||||
@register_fake("_C::awq_dequantize")
|
||||
def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
|
||||
zeros: torch.Tensor, split_k_iters: torch.SymInt,
|
||||
@ -1348,15 +1321,6 @@ def scaled_int8_quant(
|
||||
return output, input_scales, input_azp
|
||||
|
||||
|
||||
# qqq ops
|
||||
def marlin_qqq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
||||
s_tok: torch.Tensor, s_ch: torch.Tensor,
|
||||
s_group: torch.Tensor, workspace: torch.Tensor,
|
||||
size_m: int, size_n: int, size_k: int) -> torch.Tensor:
|
||||
return torch.ops._C.marlin_qqq_gemm(a, b_q_weight, s_tok, s_ch, s_group,
|
||||
workspace, size_m, size_n, size_k)
|
||||
|
||||
|
||||
# gguf
|
||||
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int, n: int,
|
||||
dtype: Optional[torch.dtype]) -> torch.Tensor:
|
||||
|
||||
@ -1112,9 +1112,9 @@ class ModelConfig:
|
||||
def _verify_quantization(self) -> None:
|
||||
supported_quantization = me_quant.QUANTIZATION_METHODS
|
||||
optimized_quantization_methods = [
|
||||
"fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
|
||||
"awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
|
||||
"quark", "modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
|
||||
"fp8", "modelopt", "gptq_marlin_24", "gptq_marlin", "awq_marlin",
|
||||
"fbgemm_fp8", "compressed-tensors", "experts_int8", "quark",
|
||||
"modelopt_fp4", "bitblas", "gptq_bitblas", "inc"
|
||||
]
|
||||
if self.quantization is not None:
|
||||
self.quantization = cast(me_quant.QuantizationMethods,
|
||||
@ -1137,7 +1137,6 @@ class ModelConfig:
|
||||
# `override_quantization_method` method) must be checked in order
|
||||
# of preference (this is particularly important for GPTQ).
|
||||
overrides = [
|
||||
"marlin",
|
||||
"bitblas",
|
||||
"gptq_marlin_24",
|
||||
"gptq_marlin",
|
||||
|
||||
@ -48,9 +48,6 @@ def _get_lora_device(base_layer: nn.Module) -> torch.device:
|
||||
# GPTQ/AWQ
|
||||
elif hasattr(base_layer, "qweight"):
|
||||
return base_layer.qweight.device
|
||||
# marlin
|
||||
elif hasattr(base_layer, "B"):
|
||||
return base_layer.B.device
|
||||
# HQQ marlin
|
||||
elif hasattr(base_layer, "W_q"):
|
||||
return base_layer.W_q.device
|
||||
|
||||
@ -42,7 +42,6 @@ WEIGHT_LOADER_V2_SUPPORTED = [
|
||||
"GPTQMarlinLinearMethod",
|
||||
"Fp8LinearMethod",
|
||||
"MarlinLinearMethod",
|
||||
"QQQLinearMethod",
|
||||
"GPTQMarlin24LinearMethod",
|
||||
"TPUInt8LinearMethod",
|
||||
"GPTQLinearMethod",
|
||||
|
||||
@ -15,7 +15,6 @@ QuantizationMethods = Literal[
|
||||
"fbgemm_fp8",
|
||||
"modelopt",
|
||||
"modelopt_fp4",
|
||||
"marlin",
|
||||
"bitblas",
|
||||
"gguf",
|
||||
"gptq_marlin_24",
|
||||
@ -25,7 +24,6 @@ QuantizationMethods = Literal[
|
||||
"gptq",
|
||||
"compressed-tensors",
|
||||
"bitsandbytes",
|
||||
"qqq",
|
||||
"hqq",
|
||||
"experts_int8",
|
||||
"neuron_quant",
|
||||
@ -106,13 +104,11 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
from .hqq_marlin import HQQMarlinConfig
|
||||
from .inc import INCConfig
|
||||
from .ipex_quant import IPEXConfig
|
||||
from .marlin import MarlinConfig
|
||||
from .modelopt import ModelOptFp8Config, ModelOptNvFp4Config
|
||||
from .moe_wna16 import MoeWNA16Config
|
||||
from .mxfp4 import Mxfp4Config
|
||||
from .neuron_quant import NeuronQuantConfig
|
||||
from .ptpc_fp8 import PTPCFp8Config
|
||||
from .qqq import QQQConfig
|
||||
from .rtn import RTNConfig
|
||||
from .torchao import TorchAOConfig
|
||||
from .tpu_int8 import Int8TpuConfig
|
||||
@ -125,7 +121,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
"fbgemm_fp8": FBGEMMFp8Config,
|
||||
"modelopt": ModelOptFp8Config,
|
||||
"modelopt_fp4": ModelOptNvFp4Config,
|
||||
"marlin": MarlinConfig,
|
||||
"bitblas": BitBLASConfig,
|
||||
"gguf": GGUFConfig,
|
||||
"gptq_marlin_24": GPTQMarlin24Config,
|
||||
@ -136,7 +131,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
"compressed-tensors": CompressedTensorsConfig,
|
||||
"bitsandbytes": BitsAndBytesConfig,
|
||||
"ptpc_fp8": PTPCFp8Config,
|
||||
"qqq": QQQConfig,
|
||||
"hqq": HQQMarlinConfig,
|
||||
"experts_int8": ExpertsInt8Config,
|
||||
"neuron_quant": NeuronQuantConfig,
|
||||
|
||||
@ -1,263 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from typing import Any, Optional
|
||||
|
||||
import torch
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
|
||||
from vllm.model_executor.parameter import (BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
GroupQuantScaleParameter,
|
||||
PackedvLLMParameter)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class MarlinConfig(QuantizationConfig):
|
||||
"""Config class for Marlin.
|
||||
|
||||
Reference: https://github.com/IST-DASLab/marlin/tree/master
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
group_size: int,
|
||||
lm_head_quantized: bool,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
|
||||
# Group size for the quantization.
|
||||
self.group_size = group_size
|
||||
self.lm_head_quantized = lm_head_quantized
|
||||
if self.group_size != 128 and self.group_size != -1:
|
||||
raise ValueError(
|
||||
"Currently, only group size 128 and -1 (channelwise) "
|
||||
"is supported for Marlin, but got group_size of "
|
||||
f"{self.group_size}")
|
||||
|
||||
# 4 Bits packed into 32 bit datatype.
|
||||
self.pack_factor = 32 // 4
|
||||
|
||||
# Tile size used by marlin kernels.
|
||||
self.tile_size = 16
|
||||
|
||||
# Min out_features dim
|
||||
self.min_n_threads = 64
|
||||
|
||||
# Min in_features dim
|
||||
self.min_k_threads = 128
|
||||
|
||||
# Max parallel problems to solve at once (improves large
|
||||
# batch performance)
|
||||
self.max_parallel = 16
|
||||
|
||||
# Permutation length used by the marlin kernels.
|
||||
self.perm_len = 1024
|
||||
|
||||
def __repr__(self) -> str:
|
||||
return (f"MarlinConfig(group_size={self.group_size}, "
|
||||
f"lm_head_quantized={self.lm_head_quantized})")
|
||||
|
||||
@classmethod
|
||||
def get_name(cls) -> QuantizationMethods:
|
||||
return "marlin"
|
||||
|
||||
@classmethod
|
||||
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
|
||||
return [torch.half]
|
||||
|
||||
@classmethod
|
||||
# Need to figure it out
|
||||
def get_min_capability(cls) -> int:
|
||||
return 80
|
||||
|
||||
@classmethod
|
||||
def get_config_filenames(cls) -> list[str]:
|
||||
return ["quantize_config.json"]
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, config: dict[str, Any]) -> "MarlinConfig":
|
||||
group_size = cls.get_from_keys(config, ["group_size"])
|
||||
lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"],
|
||||
default=False)
|
||||
return cls(group_size, lm_head_quantized)
|
||||
|
||||
@classmethod
|
||||
def override_quantization_method(
|
||||
cls, hf_quant_cfg, user_quant) -> Optional[QuantizationMethods]:
|
||||
# compat: autogptq >=0.8.0 use checkpoint_format: str
|
||||
# compat: autogptq <=0.7.1 is_marlin_format: bool
|
||||
is_marlin_format = (hf_quant_cfg.get("checkpoint_format") == "marlin"
|
||||
or hf_quant_cfg.get("is_marlin_format", False))
|
||||
|
||||
is_valid_user_quant = (user_quant is None or user_quant == "gptq"
|
||||
or user_quant == "marlin")
|
||||
|
||||
if is_marlin_format and is_valid_user_quant:
|
||||
msg = ("The model is serialized in {} format. Using {} kernel.".
|
||||
format(cls.get_name(), cls.get_name()))
|
||||
logger.info(msg)
|
||||
return cls.get_name()
|
||||
|
||||
return None
|
||||
|
||||
def get_quant_method(self, layer: torch.nn.Module,
|
||||
prefix: str) -> Optional["MarlinLinearMethod"]:
|
||||
if (isinstance(layer, LinearBase) or
|
||||
(isinstance(layer, ParallelLMHead) and self.lm_head_quantized)):
|
||||
return MarlinLinearMethod(self)
|
||||
return None
|
||||
|
||||
|
||||
class MarlinLinearMethod(LinearMethodBase):
|
||||
"""Linear method for Marlin.
|
||||
|
||||
Args:
|
||||
quant_config: The Marlin quantization config.
|
||||
"""
|
||||
|
||||
def __init__(self, quant_config: MarlinConfig):
|
||||
self.quant_config = quant_config
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
input_size_per_partition: int,
|
||||
output_partition_sizes: list[int],
|
||||
input_size: int,
|
||||
output_size: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
del output_size # Unused.
|
||||
weight_loader = extra_weight_attrs["weight_loader"]
|
||||
|
||||
if params_dtype != torch.float16:
|
||||
raise ValueError(
|
||||
f"The params dtype must be float16, but got {params_dtype}")
|
||||
|
||||
# Validate output_size_per_partition
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
if output_size_per_partition % self.quant_config.min_n_threads != 0:
|
||||
raise ValueError(
|
||||
f"Weight output_size_per_partition = "
|
||||
f"{output_size_per_partition} is not divisible by "
|
||||
f"min_n_threads = {self.quant_config.min_n_threads}.")
|
||||
if output_size_per_partition % self.quant_config.pack_factor != 0:
|
||||
raise ValueError(
|
||||
f"Weight output_size_per_partition = "
|
||||
f"{output_size_per_partition} is not divisible by "
|
||||
f"pack_factor = {self.quant_config.pack_factor}.")
|
||||
|
||||
# Validate input_size_per_partition
|
||||
if input_size_per_partition % self.quant_config.min_k_threads != 0:
|
||||
raise ValueError(
|
||||
f"Weight input_size_per_partition = "
|
||||
f"{input_size_per_partition} is not divisible by "
|
||||
f"min_k_threads = {self.quant_config.min_k_threads}.")
|
||||
if (self.quant_config.group_size != -1 and
|
||||
input_size_per_partition % self.quant_config.group_size != 0):
|
||||
raise ValueError(f"Weight input_size_per_partition = "
|
||||
f"{input_size_per_partition} is not divisible by "
|
||||
f"group_size = {self.quant_config.group_size}.")
|
||||
|
||||
# Check that we have at least 4 tiles horizontally in the shard
|
||||
num_tiles_per_perm = self.quant_config.perm_len // (
|
||||
self.quant_config.tile_size**2)
|
||||
if output_size_per_partition % num_tiles_per_perm != 0:
|
||||
raise ValueError(
|
||||
"Each permutation group must reside on the same gpu")
|
||||
|
||||
# Quantized 4Bit weights packed into Int32.
|
||||
qweight = PackedvLLMParameter(
|
||||
data=torch.empty(
|
||||
input_size_per_partition // self.quant_config.tile_size,
|
||||
output_size_per_partition * self.quant_config.tile_size //
|
||||
self.quant_config.pack_factor,
|
||||
device="cuda",
|
||||
dtype=torch.int32,
|
||||
),
|
||||
input_dim=0,
|
||||
output_dim=1,
|
||||
packed_dim=1,
|
||||
packed_factor=self.quant_config.pack_factor,
|
||||
marlin_tile_size=self.quant_config.tile_size,
|
||||
weight_loader=weight_loader)
|
||||
|
||||
# Determine if channelwise or not
|
||||
input_groups = (1 if self.quant_config.group_size == -1 else
|
||||
input_size_per_partition //
|
||||
self.quant_config.group_size)
|
||||
|
||||
weight_scale_args = {
|
||||
"data":
|
||||
torch.empty(
|
||||
input_groups,
|
||||
output_size_per_partition,
|
||||
device="cuda",
|
||||
dtype=params_dtype,
|
||||
),
|
||||
"weight_loader":
|
||||
weight_loader
|
||||
}
|
||||
if input_groups == 1:
|
||||
scales = ChannelQuantScaleParameter(output_dim=1,
|
||||
**weight_scale_args)
|
||||
else:
|
||||
scales = GroupQuantScaleParameter(output_dim=1,
|
||||
input_dim=0,
|
||||
**weight_scale_args)
|
||||
|
||||
# Allocate workspace (Used for internal locking mechanism)
|
||||
max_workspace_size = (
|
||||
output_size_per_partition //
|
||||
self.quant_config.min_n_threads) * self.quant_config.max_parallel
|
||||
|
||||
workspace = BasevLLMParameter(data=torch.zeros(max_workspace_size,
|
||||
device="cuda",
|
||||
dtype=torch.int),
|
||||
weight_loader=weight_loader)
|
||||
|
||||
layer.register_parameter("B", qweight)
|
||||
layer.register_parameter("s", scales)
|
||||
layer.register_parameter("workspace", workspace)
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
# required by torch.compile
|
||||
layer.B = Parameter(layer.B.data, requires_grad=False)
|
||||
layer.s = Parameter(layer.s.data, requires_grad=False)
|
||||
layer.workspace = Parameter(layer.workspace.data, requires_grad=False)
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
qweight = layer.B
|
||||
scales = layer.s
|
||||
workspace = layer.workspace
|
||||
|
||||
x_2d = x.view(-1, x.shape[-1])
|
||||
|
||||
size_m = x_2d.shape[0]
|
||||
size_k = x_2d.shape[1]
|
||||
size_n = scales.shape[1]
|
||||
|
||||
output_2d = ops.marlin_gemm(x_2d, qweight, scales, workspace, size_m,
|
||||
size_n, size_k)
|
||||
|
||||
output = output_2d.view(x.shape[:-1] + (output_2d.shape[1], ))
|
||||
|
||||
if bias is not None:
|
||||
output.add_(bias) # In-place add
|
||||
|
||||
return output
|
||||
@ -1,275 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from typing import Any, Optional
|
||||
|
||||
import torch
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.parameter import (BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
GroupQuantScaleParameter,
|
||||
PackedvLLMParameter)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
MARLIN_QQQ_TILE = 16
|
||||
MARLIN_QQQ_MIN_THREAD_N = 64
|
||||
MARLIN_QQQ_MIN_THREAD_K = 128
|
||||
MARLIN_QQQ_MAX_PARALLEL = 16
|
||||
|
||||
MARLIN_QQQ_SUPPORTED_NUM_BITS = [4]
|
||||
MARLIN_QQQ_SUPPORTED_GROUP_SIZES = [-1, 128]
|
||||
MARLIN_QQQ_SUPPORTED_SYM = [True]
|
||||
|
||||
|
||||
class QQQConfig(QuantizationConfig):
|
||||
"""Config class for QQQ
|
||||
|
||||
Reference: https://arxiv.org/pdf/2406.09904
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
weight_bits: int,
|
||||
group_size: int,
|
||||
is_sym: bool = True,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.weight_bits = weight_bits
|
||||
self.group_size = group_size
|
||||
self.is_sym = is_sym
|
||||
|
||||
# Verify
|
||||
if self.weight_bits not in MARLIN_QQQ_SUPPORTED_NUM_BITS:
|
||||
raise ValueError(
|
||||
f"QQQ does not support weight_bits = {self.weight_bits}. "
|
||||
f"Only weight_bits = {MARLIN_QQQ_SUPPORTED_NUM_BITS} "
|
||||
"are supported.")
|
||||
if self.group_size not in MARLIN_QQQ_SUPPORTED_GROUP_SIZES:
|
||||
raise ValueError(
|
||||
f"QQQ does not support group_size = {self.group_size}. "
|
||||
f"Only group_sizes = {MARLIN_QQQ_SUPPORTED_GROUP_SIZES} "
|
||||
"are supported.")
|
||||
if self.is_sym not in MARLIN_QQQ_SUPPORTED_SYM:
|
||||
raise ValueError(
|
||||
f"QQQ does not support is_sym = {self.is_sym}. "
|
||||
f"Only sym = {MARLIN_QQQ_SUPPORTED_SYM} are supported.")
|
||||
|
||||
# 4 Bits packed into 32 bit datatype.
|
||||
self.pack_factor = 32 // self.weight_bits
|
||||
|
||||
# Tile size used by QQQ kernels.
|
||||
self.tile_size = MARLIN_QQQ_TILE
|
||||
|
||||
# Min out_features dim
|
||||
self.min_n_threads = MARLIN_QQQ_MIN_THREAD_N
|
||||
|
||||
# Min in_features dim
|
||||
self.min_k_threads = MARLIN_QQQ_MIN_THREAD_K
|
||||
|
||||
# Max parallel problems to solve at once (improves large
|
||||
# batch performance)
|
||||
self.max_parallel = MARLIN_QQQ_MAX_PARALLEL
|
||||
|
||||
# Permutation length used by the QQQ kernels.
|
||||
self.perm_len = 1024
|
||||
|
||||
def __repr__(self) -> str:
|
||||
return "QQQConfig(weight_bits={}, group_size={})".format(
|
||||
self.weight_bits, self.group_size)
|
||||
|
||||
@classmethod
|
||||
def get_name(cls) -> QuantizationMethods:
|
||||
return "qqq"
|
||||
|
||||
@classmethod
|
||||
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
|
||||
return [torch.half]
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
return 80
|
||||
|
||||
@classmethod
|
||||
def get_config_filenames(cls) -> list[str]:
|
||||
"""List of filenames to search for in the model directory."""
|
||||
return [
|
||||
"quant_config.json",
|
||||
"quantize_config.json",
|
||||
]
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, config: dict[str, Any]) -> "QQQConfig":
|
||||
weight_bits = cls.get_from_keys(config, ["wbits"])
|
||||
group_size = cls.get_from_keys(config, ["group_size"])
|
||||
return cls(weight_bits, group_size)
|
||||
|
||||
def get_quant_method(self, layer: torch.nn.Module,
|
||||
prefix: str) -> Optional["QQQLinearMethod"]:
|
||||
if isinstance(layer, LinearBase):
|
||||
return QQQLinearMethod(self)
|
||||
return None
|
||||
|
||||
|
||||
class QQQLinearMethod(LinearMethodBase):
|
||||
"""Linear method for QQQ.
|
||||
|
||||
Args:
|
||||
quant_config: The QQQ quantization config.
|
||||
"""
|
||||
|
||||
def __init__(self, quant_config: QQQConfig):
|
||||
self.quant_config = quant_config
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
input_size_per_partition: int,
|
||||
output_partition_sizes: list[int],
|
||||
input_size: int,
|
||||
output_size: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
weight_loader = extra_weight_attrs["weight_loader"]
|
||||
if params_dtype != torch.float16:
|
||||
raise ValueError(
|
||||
f"The params dtype must be float16, but got {params_dtype}")
|
||||
|
||||
# Validate output_size_per_partition
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
if output_size_per_partition % self.quant_config.min_n_threads != 0:
|
||||
raise ValueError(
|
||||
f"Weight output_size_per_partition = "
|
||||
f"{output_size_per_partition} is not divisible by "
|
||||
f"min_n_threads = {self.quant_config.min_n_threads}.")
|
||||
if output_size_per_partition % self.quant_config.pack_factor != 0:
|
||||
raise ValueError(
|
||||
f"Weight output_size_per_partition = "
|
||||
f"{output_size_per_partition} is not divisible by "
|
||||
f"pack_factor = {self.quant_config.pack_factor}.")
|
||||
|
||||
# Validate input_size_per_partition
|
||||
if input_size_per_partition % self.quant_config.min_k_threads != 0:
|
||||
raise ValueError(
|
||||
f"Weight input_size_per_partition = "
|
||||
f"{input_size_per_partition} is not divisible by "
|
||||
f"min_k_threads = {self.quant_config.min_k_threads}.")
|
||||
if (self.quant_config.group_size != -1 and
|
||||
input_size_per_partition % self.quant_config.group_size != 0):
|
||||
raise ValueError(f"Weight input_size_per_partition = "
|
||||
f"{input_size_per_partition} is not divisible by "
|
||||
f"group_size = {self.quant_config.group_size}.")
|
||||
|
||||
# Check that we have at least 4 tiles horizontally in the shard
|
||||
num_tiles_per_perm = self.quant_config.perm_len // (
|
||||
self.quant_config.tile_size**2)
|
||||
if output_size_per_partition % num_tiles_per_perm != 0:
|
||||
raise ValueError(
|
||||
"Each permutation group must reside on the same gpu")
|
||||
|
||||
# Quantized 4Bit weights packed into Int32.
|
||||
qweight = PackedvLLMParameter(
|
||||
data=torch.empty(
|
||||
input_size_per_partition // self.quant_config.tile_size,
|
||||
output_size_per_partition * self.quant_config.tile_size //
|
||||
self.quant_config.pack_factor,
|
||||
device="cuda",
|
||||
dtype=torch.int32,
|
||||
),
|
||||
input_dim=0,
|
||||
output_dim=1,
|
||||
packed_dim=1,
|
||||
packed_factor=self.quant_config.pack_factor,
|
||||
marlin_tile_size=self.quant_config.tile_size,
|
||||
weight_loader=weight_loader)
|
||||
|
||||
s_channel = ChannelQuantScaleParameter(data=torch.empty(
|
||||
1,
|
||||
output_size_per_partition,
|
||||
device="cuda",
|
||||
dtype=torch.float,
|
||||
),
|
||||
weight_loader=weight_loader,
|
||||
output_dim=1)
|
||||
|
||||
if self.quant_config.group_size == -1:
|
||||
s_group_data = torch.tensor(
|
||||
[],
|
||||
device="cuda",
|
||||
dtype=torch.half,
|
||||
)
|
||||
else:
|
||||
s_group_data = torch.empty(
|
||||
input_size_per_partition // self.quant_config.group_size,
|
||||
output_size_per_partition,
|
||||
device="cuda",
|
||||
dtype=torch.half,
|
||||
)
|
||||
|
||||
s_group_attr = {"data": s_group_data, "weight_loader": weight_loader}
|
||||
|
||||
if self.quant_config.group_size == -1:
|
||||
s_group = BasevLLMParameter(**s_group_attr)
|
||||
else:
|
||||
s_group = GroupQuantScaleParameter(output_dim=1,
|
||||
input_dim=0,
|
||||
**s_group_attr)
|
||||
|
||||
# Allocate workspace (Used for internal locking mechanism)
|
||||
max_workspace_size = (
|
||||
output_size_per_partition //
|
||||
self.quant_config.min_n_threads) * self.quant_config.max_parallel
|
||||
|
||||
workspace = BasevLLMParameter(data=torch.zeros(max_workspace_size,
|
||||
device="cuda",
|
||||
dtype=torch.int),
|
||||
weight_loader=weight_loader)
|
||||
|
||||
layer.register_parameter("B", qweight)
|
||||
layer.register_parameter("s_channel", s_channel)
|
||||
layer.register_parameter("s_group", s_group)
|
||||
layer.register_parameter("workspace", workspace)
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
# required by torch.compile
|
||||
layer.B = Parameter(layer.B.data, requires_grad=False)
|
||||
layer.s_channel = Parameter(layer.s_channel.data, requires_grad=False)
|
||||
layer.s_group = Parameter(layer.s_group.data, requires_grad=False)
|
||||
layer.workspace = Parameter(layer.workspace.data, requires_grad=False)
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
qweight = layer.B
|
||||
s_ch = layer.s_channel
|
||||
s_group = layer.s_group
|
||||
workspace = layer.workspace
|
||||
|
||||
x_2d = x.view(-1, x.shape[-1])
|
||||
|
||||
size_m = x_2d.shape[0]
|
||||
size_k = x_2d.shape[1]
|
||||
size_n = s_ch.shape[1]
|
||||
|
||||
x_int8, s_tok, _ = ops.scaled_int8_quant(x_2d)
|
||||
|
||||
output_2d = ops.marlin_qqq_gemm(x_int8, qweight, s_tok, s_ch, s_group,
|
||||
workspace, size_m, size_n, size_k)
|
||||
|
||||
output = output_2d.view(x.shape[:-1] + (output_2d.shape[1], ))
|
||||
|
||||
if bias is not None:
|
||||
output.add_(bias) # In-place add
|
||||
|
||||
return output
|
||||
@ -1,126 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
from .marlin_utils_test import marlin_permute_weights
|
||||
from .quant_utils import get_pack_factor, qqq_quantize_weights
|
||||
|
||||
|
||||
def marlin_qqq_weights(q_w, size_k, size_n, num_bits, perm, group_size):
|
||||
# Permute
|
||||
q_w = marlin_permute_weights(q_w, size_k, size_n, perm)
|
||||
|
||||
# Pack
|
||||
pack_factor = get_pack_factor(num_bits)
|
||||
orig_device = q_w.device
|
||||
|
||||
q_w = q_w.cpu().numpy().astype(numpy.uint32)
|
||||
|
||||
q_packed = numpy.zeros((q_w.shape[0], q_w.shape[1] // pack_factor),
|
||||
dtype=numpy.uint32)
|
||||
if group_size == size_k:
|
||||
for i in range(pack_factor):
|
||||
q_packed |= (q_w[:, i::pack_factor] & 0xF) << num_bits * i
|
||||
else:
|
||||
for i in range(pack_factor):
|
||||
q_packed |= q_w[:, i::pack_factor] << num_bits * i
|
||||
|
||||
q_packed = torch.from_numpy(q_packed.astype(numpy.int32)).to(orig_device)
|
||||
|
||||
return q_packed
|
||||
|
||||
|
||||
def get_qqq_scale_perms():
|
||||
scale_perm: list[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single: list[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return scale_perm, scale_perm_single
|
||||
|
||||
|
||||
# NOTE(HandH1998): QQQ employs different perms for per-group and per-channel weight quantization. # noqa: E501
|
||||
def get_qqq_weight_perm(num_bits: int, quant_type: str):
|
||||
perm_list: list[int] = []
|
||||
for i in range(32):
|
||||
perm1: list[int] = []
|
||||
col = i // 4
|
||||
for block in [0, 1]:
|
||||
for row in [
|
||||
4 * (i % 4),
|
||||
4 * (i % 4) + 1,
|
||||
4 * (i % 4) + 2,
|
||||
4 * (i % 4) + 3,
|
||||
]:
|
||||
perm1.append(16 * row + col + 8 * block)
|
||||
for j in range(4):
|
||||
perm_list.extend([p + 256 * j for p in perm1])
|
||||
|
||||
perm = numpy.array(perm_list)
|
||||
|
||||
assert quant_type in ["per-channel",
|
||||
"per-group"], "not supported quantization type"
|
||||
if num_bits == 4:
|
||||
if quant_type == "per-channel":
|
||||
interleave = numpy.array([4, 0, 5, 1, 6, 2, 7, 3])
|
||||
else:
|
||||
interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7])
|
||||
else:
|
||||
raise Exception("num_bits must be 4, got {}".format(num_bits))
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
return perm
|
||||
|
||||
|
||||
def marlin_qqq_permute_scales(s_group, s_channel, size_k, size_n, group_size):
|
||||
scale_perm, scale_perm_single = get_qqq_scale_perms()
|
||||
if group_size < size_k and group_size != -1:
|
||||
s_group = s_group.reshape((-1, len(scale_perm)))[:, scale_perm]
|
||||
s_channel = s_channel.reshape(
|
||||
(-1, len(scale_perm_single)))[:, scale_perm_single]
|
||||
s_group = s_group.reshape((-1, size_n)).contiguous()
|
||||
else:
|
||||
s_channel = s_channel.reshape(
|
||||
(-1, len(scale_perm_single)))[:, scale_perm_single]
|
||||
s_channel = s_channel.reshape((-1, size_n)).contiguous()
|
||||
|
||||
return s_group, s_channel
|
||||
|
||||
|
||||
def marlin_qqq_quantize(
|
||||
w: torch.Tensor,
|
||||
num_bits: int,
|
||||
group_size: int,
|
||||
):
|
||||
size_k, size_n = w.shape
|
||||
|
||||
# Normalize group_size
|
||||
if group_size == -1:
|
||||
group_size = size_k
|
||||
assert group_size <= size_k
|
||||
quant_type = "per-channel" if group_size == size_k else "per-group"
|
||||
|
||||
# Quantize
|
||||
w_ref, q_w, s_group, s_channel = qqq_quantize_weights(
|
||||
w, num_bits, group_size)
|
||||
|
||||
# Reformat to marlin_qqq
|
||||
weight_perm = get_qqq_weight_perm(num_bits, quant_type)
|
||||
marlin_qqq_q_w = marlin_qqq_weights(q_w, size_k, size_n, num_bits,
|
||||
weight_perm, group_size)
|
||||
marlin_qqq_s_group, marlin_qqq_s_channel = marlin_qqq_permute_scales(
|
||||
s_group, s_channel, size_k, size_n, group_size)
|
||||
|
||||
# Create result
|
||||
res_list = [
|
||||
w_ref, marlin_qqq_q_w, marlin_qqq_s_group, marlin_qqq_s_channel
|
||||
]
|
||||
for i in range(len(res_list)):
|
||||
res_list[i] = res_list[i].to(w.device)
|
||||
|
||||
return res_list
|
||||
@ -9,8 +9,6 @@ import numpy
|
||||
import torch
|
||||
|
||||
from vllm._custom_ops import cutlass_scaled_mm_supports_fp4
|
||||
from vllm.model_executor.layers.quantization.qqq import (
|
||||
MARLIN_QQQ_SUPPORTED_NUM_BITS)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.scalar_type import ScalarType, scalar_types
|
||||
|
||||
@ -386,89 +384,6 @@ def gptq_quantize_weights(w: torch.Tensor,
|
||||
return w_ref, w_q, w_s, g_idx, rand_perm
|
||||
|
||||
|
||||
# QQQ employs different quant schemes for per-group and
|
||||
# per-channel quantization.
|
||||
def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int):
|
||||
orig_device = w.device
|
||||
size_k, size_n = w.shape
|
||||
|
||||
assert w.is_floating_point(), "w must be float"
|
||||
assert num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS, \
|
||||
f"Unsupported num_bits = {num_bits}"
|
||||
assert group_size in SUPPORTED_GROUP_SIZES + [
|
||||
size_k
|
||||
], f"Unsupported groupsize = {group_size}"
|
||||
|
||||
if group_size == -1:
|
||||
group_size = size_k
|
||||
assert group_size <= size_k
|
||||
|
||||
if group_size < size_k:
|
||||
# Reshape to [groupsize, -1]
|
||||
w = w.reshape((-1, group_size, size_n))
|
||||
w = w.permute(1, 0, 2)
|
||||
w = w.reshape((group_size, -1))
|
||||
|
||||
max_q_val = 2**num_bits - 1
|
||||
half_q_val = (max_q_val + 1) // 2
|
||||
|
||||
# Compute scale for each group
|
||||
s_group = torch.max(torch.abs(w), 0, keepdim=True)[0]
|
||||
s_group *= 2 / max_q_val # 2 => symmetric
|
||||
|
||||
# Quantize
|
||||
q_w = torch.round(w / s_group).int()
|
||||
q_w += half_q_val
|
||||
q_w = torch.clamp(q_w, 0, max_q_val)
|
||||
# Compute ref (dequantized)
|
||||
w_ref = (q_w - half_q_val).half() * s_group
|
||||
|
||||
# Restore original shapes
|
||||
def reshape_w(w):
|
||||
w = w.reshape((group_size, -1, size_n))
|
||||
w = w.permute(1, 0, 2)
|
||||
w = w.reshape((size_k, size_n)).contiguous()
|
||||
return w
|
||||
|
||||
q_w = reshape_w(q_w)
|
||||
w_ref = reshape_w(w_ref)
|
||||
|
||||
# Compute int8 quantization scale for each channel
|
||||
s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0]
|
||||
s_channel /= 127.0
|
||||
t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8)
|
||||
w_ref = t_int8.half() * s_channel
|
||||
s_channel = s_channel.reshape(1, -1).to(dtype=torch.float)
|
||||
|
||||
# Fuse scales
|
||||
s_group = (s_group.reshape(-1, size_n).contiguous() /
|
||||
s_channel).to(dtype=torch.half)
|
||||
else:
|
||||
max_q_val = 2**(num_bits - 1) - 1
|
||||
|
||||
# Compute scale for each channel
|
||||
s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0]
|
||||
s_channel /= max_q_val
|
||||
|
||||
# Quantize
|
||||
q_w = torch.round(w / s_channel).int()
|
||||
q_w = torch.clamp(q_w, -max_q_val, max_q_val)
|
||||
# Compute ref (dequantized)
|
||||
w_ref = q_w.half() * s_channel
|
||||
|
||||
s_group = torch.tensor([], dtype=torch.half)
|
||||
# div 2 ** (8 - self.bits)) to offset right shift in unpacking
|
||||
s_channel /= (2**(8 - num_bits))
|
||||
s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float)
|
||||
|
||||
return (
|
||||
w_ref.to(device=orig_device),
|
||||
q_w.to(device=orig_device),
|
||||
s_group.to(device=orig_device),
|
||||
s_channel.to(device=orig_device),
|
||||
)
|
||||
|
||||
|
||||
def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor):
|
||||
orig_device = q_w.device
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user