Skip to content

GitLab

Commit 2eefe3d6 authored by luopl's avatar luopl
Browse files

add mamba

parent b7535e7c
Pipeline #1735 failed with stages
in 0 seconds
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2440 additions and 0 deletions
mamba/.gitignore 0 → 100644
View file @ 2eefe3d6
*__pycache__/
*.egg-info/
build/
**.so
*.hip
*_hip.*
\ No newline at end of file
mamba/.gitmodules 0 → 100644
View file @ 2eefe3d6
[submodule "3rdparty/lm-evaluation-harness"]
path = 3rdparty/lm-evaluation-harness
url = https://github.com/EleutherAI/lm-evaluation-harness/
mamba/AUTHORS 0 → 100644
View file @ 2eefe3d6
Tri Dao, tri@tridao.me
Albert Gu, agu@andrew.cmu.edu
mamba/LICENSE 0 → 100644
View file @ 2eefe3d6
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 2023 Tri Dao, Albert Gu
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.
mamba/README.md 0 → 100644
View file @ 2eefe3d6
# Mamba
![Mamba](assets/selection.png "Selective State Space")
> **Mamba: Linear-Time Sequence Modeling with Selective State Spaces**\
> Albert Gu*, Tri Dao*\
> Paper: https://arxiv.org/abs/2312.00752
![Mamba-2](assets/ssd_algorithm.png "State Space Dual Model")
> **Transformers are SSMs: Generalized Models and Efficient Algorithms**\
> **Through Structured State Space Duality**\
> Tri Dao*, Albert Gu*\
> Paper: https://arxiv.org/abs/2405.21060
## About
Mamba is a new state space model architecture showing promising performance on information-dense data such as language modeling, where previous subquadratic models fall short of Transformers.
It is based on the line of progress on [structured state space models](https://github.com/state-spaces/s4),
with an efficient hardware-aware design and implementation in the spirit of [FlashAttention](https://github.com/Dao-AILab/flash-attention).
## Installation
- [Option] `pip install causal-conv1d>=1.4.0`: an efficient implementation of a simple causal Conv1d layer used inside the Mamba block.
- `pip install mamba-ssm`: the core Mamba package.
- `pip install mamba-ssm[causal-conv1d]`: To install core Mamba package and causal-conv1d.
- `pip install mamba-ssm[dev]`: To install core Mamba package and dev depdencies.
It can also be built from source with `pip install .` from this repository.
If `pip` complains about PyTorch versions, try passing `--no-build-isolation` to `pip`.
Other requirements:
- Linux
- NVIDIA GPU
- PyTorch 1.12+
- CUDA 11.6+
For AMD cards, see additional prerequisites below.
## Usage
We expose several levels of interface with the Mamba model.
### Selective SSM
Mamba is based on a selective SSM layer, which is the focus of the paper (Section 3; Algorithm 2).
Source: [ops/selective_scan_interface.py](mamba_ssm/ops/selective_scan_interface.py).
### Mamba Block
The main module of this repository is the Mamba architecture block wrapping the selective SSM.
Source: [modules/mamba_simple.py](mamba_ssm/modules/mamba_simple.py).
Usage:
``` python
import torch
from mamba_ssm import Mamba
batch, length, dim = 2, 64, 16
x = torch.randn(batch, length, dim).to("cuda")
model = Mamba(
# This module uses roughly 3 * expand * d_model^2 parameters
d_model=dim, # Model dimension d_model
d_state=16, # SSM state expansion factor
d_conv=4, # Local convolution width
expand=2, # Block expansion factor
).to("cuda")
y = model(x)
assert y.shape == x.shape
```
### Mamba-2
The Mamba-2 block is implemented at [modules/mamba2.py](mamba_ssm/modules/mamba2.py).
A simpler version is at [modules/mamba2_simple.py](mamba_ssm/modules/mamba2_simple.py)
The usage is similar to Mamba(-1):
``` python
from mamba_ssm import Mamba2
model = Mamba2(
# This module uses roughly 3 * expand * d_model^2 parameters
d_model=dim, # Model dimension d_model
d_state=64, # SSM state expansion factor, typically 64 or 128
d_conv=4, # Local convolution width
expand=2, # Block expansion factor
).to("cuda")
y = model(x)
assert y.shape == x.shape
```
#### SSD
A minimal version of the inner SSD module (Listing 1 from the Mamba-2 paper) with conversion between "discrete" and "continuous" SSM versions
is at [modules/ssd_minimal.py](mamba_ssm/modules/ssd_minimal.py).
### Mamba Language Model
Finally, we provide an example of a complete language model: a deep sequence model backbone (with repeating Mamba blocks) + language model head.
Source: [models/mixer_seq_simple.py](mamba_ssm/models/mixer_seq_simple.py).
This is an example of how to integrate Mamba into an end-to-end neural network.
This example is used in the generation scripts below.
## Pretrained Models
Pretrained models are uploaded to
[Hugging Face](https://huggingface.co/state-spaces): `mamba-130m`, `mamba-370m`,
`mamba-790m`, `mamba-1.4b`, `mamba-2.8b`, `mamba2-130m`, `mamba2-370m`,
`mamba2-780m`, `mamba2-1.3b`, `mamba2-2.7b`, `transformerpp-2.7b`, `mamba2attn-2.7b`, trained on 300B tokens on the Pile, as well as `mamba-2.8b-slimpj`
(trained on 600B tokens on the SlimPajama dataset).
The models will be autodownloaded by the generation script below.
These models were trained on the [Pile](https://huggingface.co/datasets/EleutherAI/pile), and follow the standard model dimensions described by GPT-3 and followed by many open source models:
| Parameters | Layers | Model dim. |
|------------|--------|------------|
| 130M | 24 | 768 |
| 370M | 48 | 1024 |
| 790M | 48 | 1536 |
| 1.4B | 48 | 2048 |
| 2.8B | 64 | 2560 |
(The layer count of Mamba doubles that of a Transformer with similar size, as two Mamba blocks are needed for each "layer" (MHA block + MLP block) of a Transformer.)
Note: these are base models trained only for 300B tokens, without any form of downstream modification (instruction tuning, etc.).
Performance is expected to be comparable or better than other architectures trained on similar data, but not to match larger or fine-tuned models.
## Evaluations
To run zero-shot evaluations of models (corresponding to Table 3 of the paper),
we use the
[lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)
library.
1. Install `lm-evaluation-harness` by `pip install lm-eval==0.4.2`.
2. Run evaluation with (more documentation at the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor) repo):
``` sh
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/mamba-130m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa --device cuda --batch_size 256
python evals/lm_harness_eval.py --model hf --model_args pretrained=EleutherAI/pythia-160m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande --device cuda --batch_size 64
```
To reproduce the results on the `mamba-2.8b-slimpj` model reported in the blogposts:
``` sh
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/mamba-2.8b-slimpj --tasks boolq,piqa,hellaswag,winogrande,arc_easy,arc_challenge,openbookqa,race,truthfulqa_mc2 --device cuda --batch_size 256
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/mamba-2.8b-slimpj --tasks mmlu --num_fewshot 5 --device cuda --batch_size 256
```
To run evaluations on Mamba-2 models, simply replace the model names:
``` sh
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/mamba2-2.7b --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa --device cuda --batch_size 256
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/transformerpp-2.7b --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa --device cuda --batch_size 256
lm_eval --model mamba_ssm --model_args pretrained=state-spaces/mamba2attn-2.7b --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa --device cuda --batch_size 256
```
Note that the result of each task might differ from reported values by 0.1-0.3 due to noise in the evaluation process.
## Inference
The script [benchmarks/benchmark_generation_mamba_simple.py](benchmarks/benchmark_generation_mamba_simple.py)
1. autoloads a model from the Hugging Face Hub,
2. generates completions of a user-specified prompt,
3. benchmarks the inference speed of this generation.
Other configurable options include the top-p (nucleus sampling) probability, and the softmax temperature.
### Examples
To test generation latency (e.g. batch size = 1) with different sampling strategies:
``` sh
python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.7 --repetition-penalty 1.2
python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.7 --repetition-penalty 1.2
python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --minp 0.05 --topk 0 --temperature 0.7 --repetition-penalty 1.2
```
To test generation throughput with random prompts (e.g. large batch size):
``` sh
python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --batch 64
python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --batch 64
```
With Mamba-2, you just need to change the model name:
``` sh
python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba2-2.7b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.7 --repetition-penalty 1.2
```
## Troubleshooting
### Precision
Our models were trained using PyTorch [AMP](https://pytorch.org/docs/stable/amp.html) for mixed precision. AMP keeps model parameters in float32 and casts to half precision when necessary.
On the other hand, other frameworks like DeepSpeed store parameters in float16 and upcasts when necessary (e.g. for optimizer accumulation).
We've observed that higher precision for the main model parameters may be necessary, because SSMs are sensitive to their recurrent dynamics. If you are experiencing instabilities,
as a first step please try a framework storing parameters in fp32 (such as AMP).
### Initialization
Some parts of the model have initializations inherited from prior work on S4 models.
For [example](https://github.com/state-spaces/mamba/blob/f0affcf69f06d1d06cef018ff640bf080a11c421/mamba_ssm/modules/mamba_simple.py#L102), the $\Delta$ parameter has a targeted range by initializing the bias of its linear projection.
However, some frameworks may have post-initialization hooks (e.g. setting all bias terms in `nn.Linear` modules to zero).
If this is the case, you may have to add custom logic (e.g. this [line](https://github.com/state-spaces/mamba/blob/f0affcf69f06d1d06cef018ff640bf080a11c421/mamba_ssm/modules/mamba_simple.py#L104) turns off re-initializing in our trainer, but would be a no-op in any other framework)
that is specific to the training framework.
## Additional Prerequisites for AMD cards
### Patching ROCm
If you are on ROCm 6.0, run the following steps to avoid errors during compilation. This is not required for ROCm 6.1 onwards.
1. Locate your ROCm installation directory. This is typically found at `/opt/rocm/`, but may vary depending on your installation.
2. Apply the Patch. Run with `sudo` in case you encounter permission issues.
```bash
patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h < rocm_patch/rocm6_0.patch
```
## Citation
If you use this codebase, or otherwise find our work valuable, please cite Mamba:
```
@article{mamba,
title={Mamba: Linear-Time Sequence Modeling with Selective State Spaces},
author={Gu, Albert and Dao, Tri},
journal={arXiv preprint arXiv:2312.00752},
year={2023}
}
@inproceedings{mamba2,
title={Transformers are {SSM}s: Generalized Models and Efficient Algorithms Through Structured State Space Duality},
author={Dao, Tri and Gu, Albert},
booktitle={International Conference on Machine Learning (ICML)},
year={2024}
}
```
mamba/benchmarks/benchmark_generation_mamba_simple.py 0 → 100644
View file @ 2eefe3d6
# Copyright (c) 2023, Tri Dao, Albert Gu.
import argparse
import time
import json
import torch
import torch.nn.functional as F
from einops import rearrange
from transformers import AutoTokenizer, AutoModelForCausalLM
from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
parser = argparse.ArgumentParser(description="Generation benchmarking")
parser.add_argument("--model-name", type=str, default="state-spaces/mamba-130m")
parser.add_argument("--prompt", type=str, default=None)
parser.add_argument("--promptlen", type=int, default=100)
parser.add_argument("--genlen", type=int, default=100)
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--topk", type=int, default=1)
parser.add_argument("--topp", type=float, default=1.0)
parser.add_argument("--minp", type=float, default=0.0)
parser.add_argument("--repetition-penalty", type=float, default=1.0)
parser.add_argument("--batch", type=int, default=1)
args = parser.parse_args()
repeats = 3
device = "cuda"
dtype = torch.float16
print(f"Loading model {args.model_name}")
is_mamba = args.model_name.startswith("state-spaces/mamba") or args.model_name.startswith("state-spaces/transformerpp")
if is_mamba:
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
model = MambaLMHeadModel.from_pretrained(args.model_name, device=device, dtype=dtype)
else:
tokenizer = AutoTokenizer.from_pretrained(args.model_name)
model = AutoModelForCausalLM.from_pretrained(args.model_name, device_map={"": device}, torch_dtype=dtype)
model.eval()
print(f"Number of parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
torch.random.manual_seed(0)
if args.prompt is None:
input_ids = torch.randint(1, 1000, (args.batch, args.promptlen), dtype=torch.long, device="cuda")
attn_mask = torch.ones_like(input_ids, dtype=torch.long, device="cuda")
else:
tokens = tokenizer(args.prompt, return_tensors="pt")
input_ids = tokens.input_ids.to(device=device)
attn_mask = tokens.attention_mask.to(device=device)
max_length = input_ids.shape[1] + args.genlen
if is_mamba:
fn = lambda: model.generate(
input_ids=input_ids,
max_length=max_length,
cg=True,
return_dict_in_generate=True,
output_scores=True,
enable_timing=False,
temperature=args.temperature,
top_k=args.topk,
top_p=args.topp,
min_p=args.minp,
repetition_penalty=args.repetition_penalty,
)
else:
fn = lambda: model.generate(
input_ids=input_ids,
attention_mask=attn_mask,
max_length=max_length,
return_dict_in_generate=True,
pad_token_id=tokenizer.eos_token_id,
do_sample=True,
temperature=args.temperature,
top_k=args.topk,
top_p=args.topp,
repetition_penalty=args.repetition_penalty,
)
out = fn()
if args.prompt is not None:
print(tokenizer.batch_decode(out.sequences.tolist()))
torch.cuda.synchronize()
start = time.time()
for _ in range(repeats):
fn()
torch.cuda.synchronize()
print(f"Prompt length: {len(input_ids[0])}, generation length: {len(out.sequences[0]) - len(input_ids[0])}")
print(f"{args.model_name} prompt processing + decoding time: {(time.time() - start) / repeats * 1000:.0f}ms")
mamba/csrc/selective_scan/reverse_scan.cuh 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
#pragma once
#ifndef USE_ROCM
#include <cub/config.cuh>
#include <cub/util_ptx.cuh>
#include <cub/util_type.cuh>
#include <cub/block/block_raking_layout.cuh>
// #include <cub/detail/uninitialized_copy.cuh>
#else
#include <hipcub/hipcub.hpp>
namespace cub = hipcub;
#endif
#include "uninitialized_copy.cuh"
/**
* Perform a reverse sequential reduction over \p LENGTH elements of the \p input array. The aggregate is returned.
*/
template <
int LENGTH,
typename T,
typename ReductionOp>
__device__ __forceinline__ T ThreadReverseReduce(const T (&input)[LENGTH], ReductionOp reduction_op) {
static_assert(LENGTH > 0);
T retval = input[LENGTH - 1];
#pragma unroll
for (int i = LENGTH - 2; i >= 0; --i) { retval = reduction_op(retval, input[i]); }
return retval;
}
/**
* Perform a sequential inclusive postfix reverse scan over the statically-sized \p input array, seeded with the specified \p postfix. The aggregate is returned.
*/
template <
int LENGTH,
typename T,
typename ScanOp>
__device__ __forceinline__ T ThreadReverseScanInclusive(
const T (&input)[LENGTH],
T (&output)[LENGTH],
ScanOp scan_op,
const T postfix)
{
T inclusive = postfix;
#pragma unroll
for (int i = LENGTH - 1; i >= 0; --i) {
inclusive = scan_op(inclusive, input[i]);
output[i] = inclusive;
}
return inclusive;
}
/**
* Perform a sequential exclusive postfix reverse scan over the statically-sized \p input array, seeded with the specified \p postfix. The aggregate is returned.
*/
template <
int LENGTH,
typename T,
typename ScanOp>
__device__ __forceinline__ T ThreadReverseScanExclusive(
const T (&input)[LENGTH],
T (&output)[LENGTH],
ScanOp scan_op,
const T postfix)
{
// Careful, output maybe be aliased to input
T exclusive = postfix;
T inclusive;
#pragma unroll
for (int i = LENGTH - 1; i >= 0; --i) {
inclusive = scan_op(exclusive, input[i]);
output[i] = exclusive;
exclusive = inclusive;
}
return inclusive;
}
/**
* \brief WarpReverseScan provides SHFL-based variants of parallel postfix scan of items partitioned across a CUDA thread warp.
*
* LOGICAL_WARP_THREADS must be a power-of-two
*/
template <
typename T, ///< Data type being scanned
int LOGICAL_WARP_THREADS ///< Number of threads per logical warp
>
struct WarpReverseScan {
//---------------------------------------------------------------------
// Constants and type definitions
//---------------------------------------------------------------------
/// Whether the logical warp size and the PTX warp size coincide
// In hipcub, warp_threads is defined as HIPCUB_WARP_THREADS ::rocprim::warp_size()
// While in cub, it's defined as a macro that takes a redundant unused argument.
#ifndef USE_ROCM
#define WARP_THREADS CUB_WARP_THREADS(0)
#else
#define WARP_THREADS HIPCUB_WARP_THREADS
#endif
static constexpr bool IS_ARCH_WARP = (LOGICAL_WARP_THREADS == WARP_THREADS);
/// The number of warp scan steps
static constexpr int STEPS = cub::Log2<LOGICAL_WARP_THREADS>::VALUE;
static_assert(LOGICAL_WARP_THREADS == 1 << STEPS);
//---------------------------------------------------------------------
// Thread fields
//---------------------------------------------------------------------
/// Lane index in logical warp
unsigned int lane_id;
/// Logical warp index in 32-thread physical warp
unsigned int warp_id;
/// 32-thread physical warp member mask of logical warp
unsigned int member_mask;
//---------------------------------------------------------------------
// Construction
//---------------------------------------------------------------------
/// Constructor
explicit __device__ __forceinline__
WarpReverseScan()
: lane_id(cub::LaneId())
, warp_id(IS_ARCH_WARP ? 0 : (lane_id / LOGICAL_WARP_THREADS))
, member_mask(cub::WarpMask<LOGICAL_WARP_THREADS>(warp_id))
{
if (!IS_ARCH_WARP) {
lane_id = lane_id % LOGICAL_WARP_THREADS;
}
}
/// Broadcast
__device__ __forceinline__ T Broadcast(
T input, ///< [in] The value to broadcast
int src_lane) ///< [in] Which warp lane is to do the broadcasting
{
return cub::ShuffleIndex<LOGICAL_WARP_THREADS>(input, src_lane, member_mask);
}
/// Inclusive scan
template <typename ScanOpT>
__device__ __forceinline__ void InclusiveReverseScan(
T input, ///< [in] Calling thread's input item.
T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \p input.
ScanOpT scan_op) ///< [in] Binary scan operator
{
inclusive_output = input;
#pragma unroll
for (int STEP = 0; STEP < STEPS; STEP++) {
int offset = 1 << STEP;
T temp = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
inclusive_output, offset, LOGICAL_WARP_THREADS - 1, member_mask
);
// Perform scan op if from a valid peer
inclusive_output = static_cast<int>(lane_id) >= LOGICAL_WARP_THREADS - offset
? inclusive_output : scan_op(temp, inclusive_output);
}
}
/// Exclusive scan
// Get exclusive from inclusive
template <typename ScanOpT>
__device__ __forceinline__ void ExclusiveReverseScan(
T input, ///< [in] Calling thread's input item.
T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \p input.
ScanOpT scan_op, ///< [in] Binary scan operator
T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.
{
T inclusive_output;
InclusiveReverseScan(input, inclusive_output, scan_op);
warp_aggregate = cub::ShuffleIndex<LOGICAL_WARP_THREADS>(inclusive_output, 0, member_mask);
// initial value unknown
exclusive_output = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
inclusive_output, 1, LOGICAL_WARP_THREADS - 1, member_mask
);
}
/**
* \brief Computes both inclusive and exclusive reverse scans using the specified binary scan functor across the calling warp. Because no initial value is supplied, the \p exclusive_output computed for the last <em>warp-lane</em> is undefined.
*/
template <typename ScanOpT>
__device__ __forceinline__ void ReverseScan(
T input, ///< [in] Calling thread's input item.
T &inclusive_output, ///< [out] Calling thread's inclusive-scan output item.
T &exclusive_output, ///< [out] Calling thread's exclusive-scan output item.
ScanOpT scan_op) ///< [in] Binary scan operator
{
InclusiveReverseScan(input, inclusive_output, scan_op);
// initial value unknown
exclusive_output = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
inclusive_output, 1, LOGICAL_WARP_THREADS - 1, member_mask
);
}
};
/**
* \brief BlockReverseScan provides variants of raking-based parallel postfix scan across a CUDA thread block.
*/
template <
typename T, ///< Data type being scanned
int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension
bool MEMOIZE=false ///< Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure
>
struct BlockReverseScan {
//---------------------------------------------------------------------
// Types and constants
//---------------------------------------------------------------------
/// Constants
/// The thread block size in threads
static constexpr int BLOCK_THREADS = BLOCK_DIM_X;
/// Layout type for padded thread block raking grid
using BlockRakingLayout = cub::BlockRakingLayout<T, BLOCK_THREADS>;
// The number of reduction elements is not a multiple of the number of raking threads for now
static_assert(BlockRakingLayout::UNGUARDED);
/// Number of raking threads
static constexpr int RAKING_THREADS = BlockRakingLayout::RAKING_THREADS;
/// Number of raking elements per warp synchronous raking thread
static constexpr int SEGMENT_LENGTH = BlockRakingLayout::SEGMENT_LENGTH;
/// Cooperative work can be entirely warp synchronous
static constexpr bool WARP_SYNCHRONOUS = (int(BLOCK_THREADS) == int(RAKING_THREADS));
/// WarpReverseScan utility type
using WarpReverseScan = WarpReverseScan<T, RAKING_THREADS>;
/// Shared memory storage layout type
struct _TempStorage {
typename BlockRakingLayout::TempStorage raking_grid; ///< Padded thread block raking grid
};
/// Alias wrapper allowing storage to be unioned
struct TempStorage : cub::Uninitialized<_TempStorage> {};
//---------------------------------------------------------------------
// Per-thread fields
//---------------------------------------------------------------------
// Thread fields
_TempStorage &temp_storage;
unsigned int linear_tid;
T cached_segment[SEGMENT_LENGTH];
//---------------------------------------------------------------------
// Utility methods
//---------------------------------------------------------------------
/// Performs upsweep raking reduction, returning the aggregate
template <typename ScanOp>
__device__ __forceinline__ T Upsweep(ScanOp scan_op) {
T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
// Read data into registers
#pragma unroll
for (int i = 0; i < SEGMENT_LENGTH; ++i) { cached_segment[i] = smem_raking_ptr[i]; }
T raking_partial = cached_segment[SEGMENT_LENGTH - 1];
#pragma unroll
for (int i = SEGMENT_LENGTH - 2; i >= 0; --i) {
raking_partial = scan_op(raking_partial, cached_segment[i]);
}
return raking_partial;
}
/// Performs exclusive downsweep raking scan
template <typename ScanOp>
__device__ __forceinline__ void ExclusiveDownsweep(
ScanOp scan_op,
T raking_partial)
{
T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
// Read data back into registers
if (!MEMOIZE) {
#pragma unroll
for (int i = 0; i < SEGMENT_LENGTH; ++i) { cached_segment[i] = smem_raking_ptr[i]; }
}
ThreadReverseScanExclusive(cached_segment, cached_segment, scan_op, raking_partial);
// Write data back to smem
#pragma unroll
for (int i = 0; i < SEGMENT_LENGTH; ++i) { smem_raking_ptr[i] = cached_segment[i]; }
}
//---------------------------------------------------------------------
// Constructors
//---------------------------------------------------------------------
/// Constructor
__device__ __forceinline__ BlockReverseScan(
TempStorage &temp_storage)
:
temp_storage(temp_storage.Alias()),
linear_tid(cub::RowMajorTid(BLOCK_DIM_X, 1, 1))
{}
/// Computes an exclusive thread block-wide postfix scan using the specified binary \p scan_op functor. Each thread contributes one input element. the call-back functor \p block_postfix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically postfixes the thread block's scan inputs. Also provides every thread with the block-wide \p block_aggregate of all inputs.
template <
typename ScanOp,
typename BlockPostfixCallbackOp>
__device__ __forceinline__ void ExclusiveReverseScan(
T input, ///< [in] Calling thread's input item
T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan operator
BlockPostfixCallbackOp &block_postfix_callback_op) ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a thread block-wide postfix to be applied to all inputs.
{
if (WARP_SYNCHRONOUS) {
// Short-circuit directly to warp-synchronous scan
T block_aggregate;
WarpReverseScan warp_scan;
warp_scan.ExclusiveReverseScan(input, exclusive_output, scan_op, block_aggregate);
// Obtain warp-wide postfix in lane0, then broadcast to other lanes
T block_postfix = block_postfix_callback_op(block_aggregate);
block_postfix = warp_scan.Broadcast(block_postfix, 0);
exclusive_output = linear_tid == BLOCK_THREADS - 1 ? block_postfix : scan_op(block_postfix, exclusive_output);
} else {
// Place thread partial into shared memory raking grid
T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
detail::uninitialized_copy(placement_ptr, input);
cub::CTA_SYNC();
// Reduce parallelism down to just raking threads
if (linear_tid < RAKING_THREADS) {
WarpReverseScan warp_scan;
// Raking upsweep reduction across shared partials
T upsweep_partial = Upsweep(scan_op);
// Warp-synchronous scan
T exclusive_partial, block_aggregate;
warp_scan.ExclusiveReverseScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);
// Obtain block-wide postfix in lane0, then broadcast to other lanes
T block_postfix = block_postfix_callback_op(block_aggregate);
block_postfix = warp_scan.Broadcast(block_postfix, 0);
// Update postfix with warpscan exclusive partial
T downsweep_postfix = linear_tid == RAKING_THREADS - 1
? block_postfix : scan_op(block_postfix, exclusive_partial);
// Exclusive raking downsweep scan
ExclusiveDownsweep(scan_op, downsweep_postfix);
}
cub::CTA_SYNC();
// Grab thread postfix from shared memory
exclusive_output = *placement_ptr;
// // Compute warp scan in each warp.
// // The exclusive output from the last lane in each warp is invalid.
// T inclusive_output;
// WarpReverseScan warp_scan;
// warp_scan.ReverseScan(input, inclusive_output, exclusive_output, scan_op);
// // Compute the warp-wide postfix and block-wide aggregate for each warp. Warp postfix for the last warp is invalid.
// T block_aggregate;
// T warp_postfix = ComputeWarpPostfix(scan_op, inclusive_output, block_aggregate);
// // Apply warp postfix to our lane's partial
// if (warp_id != 0) {
// exclusive_output = scan_op(warp_postfix, exclusive_output);
// if (lane_id == 0) { exclusive_output = warp_postfix; }
// }
// // Use the first warp to determine the thread block postfix, returning the result in lane0
// if (warp_id == 0) {
// T block_postfix = block_postfix_callback_op(block_aggregate);
// if (lane_id == 0) {
// // Share the postfix with all threads
// detail::uninitialized_copy(&temp_storage.block_postfix,
// block_postfix);
// exclusive_output = block_postfix; // The block postfix is the exclusive output for tid0
// }
// }
// cub::CTA_SYNC();
// // Incorporate thread block postfix into outputs
// T block_postfix = temp_storage.block_postfix;
// if (linear_tid > 0) { exclusive_output = scan_op(block_postfix, exclusive_output); }
}
}
/**
* \brief Computes an inclusive block-wide postfix scan using the specified binary \p scan_op functor. Each thread contributes an array of consecutive input elements. the call-back functor \p block_postfix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically postfixes the thread block's scan inputs. Also provides every thread with the block-wide \p block_aggregate of all inputs.
*/
template <
int ITEMS_PER_THREAD,
typename ScanOp,
typename BlockPostfixCallbackOp>
__device__ __forceinline__ void InclusiveReverseScan(
T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items
T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \p input)
ScanOp scan_op, ///< [in] Binary scan functor
BlockPostfixCallbackOp &block_postfix_callback_op) ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a block-wide postfix to be applied to the logical input sequence.
{
// Reduce consecutive thread items in registers
T thread_postfix = ThreadReverseReduce(input, scan_op);
// Exclusive thread block-scan
ExclusiveReverseScan(thread_postfix, thread_postfix, scan_op, block_postfix_callback_op);
// Inclusive scan in registers with postfix as seed
ThreadReverseScanInclusive(input, output, scan_op, thread_postfix);
}
};
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan.cpp 0 → 100644
View file @ 2eefe3d6
This diff is collapsed.
mamba/csrc/selective_scan/selective_scan.h 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
#pragma once
////////////////////////////////////////////////////////////////////////////////////////////////////
struct SSMScanParamsBase {
using index_t = uint32_t;
int batch, seqlen, n_chunks;
index_t a_batch_stride;
index_t b_batch_stride;
index_t out_batch_stride;
// Common data pointers.
void *__restrict__ a_ptr;
void *__restrict__ b_ptr;
void *__restrict__ out_ptr;
void *__restrict__ x_ptr;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
struct SSMParamsBase {
using index_t = uint32_t;
int batch, dim, seqlen, dstate, n_groups, n_chunks;
int dim_ngroups_ratio;
bool is_variable_B;
bool is_variable_C;
bool delta_softplus;
index_t A_d_stride;
index_t A_dstate_stride;
index_t B_batch_stride;
index_t B_d_stride;
index_t B_dstate_stride;
index_t B_group_stride;
index_t C_batch_stride;
index_t C_d_stride;
index_t C_dstate_stride;
index_t C_group_stride;
index_t u_batch_stride;
index_t u_d_stride;
index_t delta_batch_stride;
index_t delta_d_stride;
index_t z_batch_stride;
index_t z_d_stride;
index_t out_batch_stride;
index_t out_d_stride;
index_t out_z_batch_stride;
index_t out_z_d_stride;
// Common data pointers.
void *__restrict__ A_ptr;
void *__restrict__ B_ptr;
void *__restrict__ C_ptr;
void *__restrict__ D_ptr;
void *__restrict__ u_ptr;
void *__restrict__ delta_ptr;
void *__restrict__ delta_bias_ptr;
void *__restrict__ out_ptr;
void *__restrict__ x_ptr;
void *__restrict__ z_ptr;
void *__restrict__ out_z_ptr;
};
struct SSMParamsBwd: public SSMParamsBase {
index_t dout_batch_stride;
index_t dout_d_stride;
index_t dA_d_stride;
index_t dA_dstate_stride;
index_t dB_batch_stride;
index_t dB_group_stride;
index_t dB_d_stride;
index_t dB_dstate_stride;
index_t dC_batch_stride;
index_t dC_group_stride;
index_t dC_d_stride;
index_t dC_dstate_stride;
index_t du_batch_stride;
index_t du_d_stride;
index_t dz_batch_stride;
index_t dz_d_stride;
index_t ddelta_batch_stride;
index_t ddelta_d_stride;
// Common data pointers.
void *__restrict__ dout_ptr;
void *__restrict__ dA_ptr;
void *__restrict__ dB_ptr;
void *__restrict__ dC_ptr;
void *__restrict__ dD_ptr;
void *__restrict__ du_ptr;
void *__restrict__ dz_ptr;
void *__restrict__ ddelta_ptr;
void *__restrict__ ddelta_bias_ptr;
};
mamba/csrc/selective_scan/selective_scan_bwd_bf16_complex.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<at::BFloat16, complex_t>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_bf16_real.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<at::BFloat16, float>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_fp16_complex.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<at::Half, complex_t>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_fp16_real.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<at::Half, float>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_fp32_complex.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<float, complex_t>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_fp32_real.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_bwd_kernel.cuh"
template void selective_scan_bwd_cuda<float, float>(SSMParamsBwd &params, cudaStream_t stream);
\ No newline at end of file
mamba/csrc/selective_scan/selective_scan_bwd_kernel.cuh 0 → 100644
View file @ 2eefe3d6
This diff is collapsed.
mamba/csrc/selective_scan/selective_scan_common.h 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
#pragma once
#ifndef USE_ROCM
#include <cuda_bf16.h>
#else
#include <hip/hip_bf16.h>
#endif
#include <cuda_fp16.h>
#include <c10/util/complex.h> // For scalar_value_type
#ifndef USE_ROCM
constexpr size_t custom_max(std::initializer_list<size_t> ilist)
{
return std::max(ilist);
}
template<typename T>
constexpr T constexpr_min(T a, T b) {
return std::min(a, b);
}
#else
constexpr size_t custom_max(std::initializer_list<size_t> ilist)
{
return *std::max_element(ilist.begin(), ilist.end());
}
template<typename T>
constexpr T constexpr_min(T a, T b) {
return a < b ? a : b;
}
#endif
#define MAX_DSTATE 256
using complex_t = c10::complex<float>;
inline __device__ float2 operator+(const float2 & a, const float2 & b){
return {a.x + b.x, a.y + b.y};
}
inline __device__ float3 operator+(const float3 &a, const float3 &b) {
return {a.x + b.x, a.y + b.y, a.z + b.z};
}
inline __device__ float4 operator+(const float4 & a, const float4 & b){
return {a.x + b.x, a.y + b.y, a.z + b.z, a.w + b.w};
}
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int BYTES> struct BytesToType {};
template<> struct BytesToType<16> {
using Type = uint4;
static_assert(sizeof(Type) == 16);
};
template<> struct BytesToType<8> {
using Type = uint64_t;
static_assert(sizeof(Type) == 8);
};
template<> struct BytesToType<4> {
using Type = uint32_t;
static_assert(sizeof(Type) == 4);
};
template<> struct BytesToType<2> {
using Type = uint16_t;
static_assert(sizeof(Type) == 2);
};
template<> struct BytesToType<1> {
using Type = uint8_t;
static_assert(sizeof(Type) == 1);
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template<typename scalar_t, int N>
struct Converter{
static inline __device__ void to_float(const scalar_t (&src)[N], float (&dst)[N]) {
#pragma unroll
for (int i = 0; i < N; ++i) { dst[i] = src[i]; }
}
};
template<int N>
struct Converter<at::Half, N>{
static inline __device__ void to_float(const at::Half (&src)[N], float (&dst)[N]) {
static_assert(N % 2 == 0);
auto &src2 = reinterpret_cast<const half2 (&)[N / 2]>(src);
auto &dst2 = reinterpret_cast<float2 (&)[N / 2]>(dst);
#pragma unroll
for (int i = 0; i < N / 2; ++i) { dst2[i] = __half22float2(src2[i]); }
}
};
#if __CUDA_ARCH__ >= 800
template<int N>
struct Converter<at::BFloat16, N>{
static inline __device__ void to_float(const at::BFloat16 (&src)[N], float (&dst)[N]) {
static_assert(N % 2 == 0);
auto &src2 = reinterpret_cast<const nv_bfloat162 (&)[N / 2]>(src);
auto &dst2 = reinterpret_cast<float2 (&)[N / 2]>(dst);
#pragma unroll
for (int i = 0; i < N / 2; ++i) { dst2[i] = __bfloat1622float2(src2[i]); }
}
};
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// From https://stackoverflow.com/questions/9860711/cucomplex-h-and-exp
// and https://forums.developer.nvidia.com/t/complex-number-exponential-function/24696
__device__ __forceinline__ complex_t cexp2f(complex_t z) {
float t = exp2f(z.real_);
float c, s;
sincosf(z.imag_, &s, &c);
return complex_t(c * t, s * t);
}
__device__ __forceinline__ complex_t cexpf(complex_t z) {
float t = expf(z.real_);
float c, s;
sincosf(z.imag_, &s, &c);
return complex_t(c * t, s * t);
}
template<typename scalar_t> struct SSMScanOp;
template<>
struct SSMScanOp<float> {
__device__ __forceinline__ float2 operator()(const float2 &ab0, const float2 &ab1) const {
return make_float2(ab1.x * ab0.x, ab1.x * ab0.y + ab1.y);
}
};
template<>
struct SSMScanOp<complex_t> {
__device__ __forceinline__ float4 operator()(const float4 &ab0, const float4 &ab1) const {
complex_t a0 = complex_t(ab0.x, ab0.y);
complex_t b0 = complex_t(ab0.z, ab0.w);
complex_t a1 = complex_t(ab1.x, ab1.y);
complex_t b1 = complex_t(ab1.z, ab1.w);
complex_t out_a = a1 * a0;
complex_t out_b = a1 * b0 + b1;
return make_float4(out_a.real_, out_a.imag_, out_b.real_, out_b.imag_);
}
};
// A stateful callback functor that maintains a running prefix to be applied
// during consecutive scan operations.
template <typename scalar_t> struct SSMScanPrefixCallbackOp {
using scan_t = std::conditional_t<std::is_same_v<scalar_t, float>, float2, float4>;
scan_t running_prefix;
// Constructor
__device__ SSMScanPrefixCallbackOp(scan_t running_prefix_) : running_prefix(running_prefix_) {}
// Callback operator to be entered by the first warp of threads in the block.
// Thread-0 is responsible for returning a value for seeding the block-wide scan.
__device__ scan_t operator()(scan_t block_aggregate) {
scan_t old_prefix = running_prefix;
running_prefix = SSMScanOp<scalar_t>()(running_prefix, block_aggregate);
return old_prefix;
}
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template<typename Ktraits>
inline __device__ void load_input(typename Ktraits::input_t *u,
typename Ktraits::input_t (&u_vals)[Ktraits::kNItems],
typename Ktraits::BlockLoadT::TempStorage &smem_load,
int seqlen) {
if constexpr (Ktraits::kIsEvenLen) {
auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_load);
using vec_t = typename Ktraits::vec_t;
typename Ktraits::BlockLoadVecT(smem_load_vec).Load(
reinterpret_cast<vec_t*>(u),
reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(u_vals)
#ifdef USE_ROCM
, Ktraits::kNThreads * Ktraits::kNLoads
#endif
);
} else {
typename Ktraits::BlockLoadT(smem_load).Load(u, u_vals, seqlen, 0.f);
}
}
template<typename Ktraits>
inline __device__ void load_weight(typename Ktraits::input_t *Bvar,
typename Ktraits::weight_t (&B_vals)[Ktraits::kNItems],
typename Ktraits::BlockLoadWeightT::TempStorage &smem_load_weight,
int seqlen) {
constexpr int kNItems = Ktraits::kNItems;
if constexpr (!Ktraits::kIsComplex) {
typename Ktraits::input_t B_vals_load[kNItems];
if constexpr (Ktraits::kIsEvenLen) {
auto& smem_load_weight_vec = reinterpret_cast<typename Ktraits::BlockLoadWeightVecT::TempStorage&>(smem_load_weight);
using vec_t = typename Ktraits::vec_t;
typename Ktraits::BlockLoadWeightVecT(smem_load_weight_vec).Load(
reinterpret_cast<vec_t*>(Bvar),
reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(B_vals_load)
);
} else {
typename Ktraits::BlockLoadWeightT(smem_load_weight).Load(Bvar, B_vals_load, seqlen, 0.f);
}
// #pragma unroll
// for (int i = 0; i < kNItems; ++i) { B_vals[i] = B_vals_load[i]; }
Converter<typename Ktraits::input_t, kNItems>::to_float(B_vals_load, B_vals);
} else {
typename Ktraits::input_t B_vals_load[kNItems * 2];
if constexpr (Ktraits::kIsEvenLen) {
auto& smem_load_weight_vec = reinterpret_cast<typename Ktraits::BlockLoadWeightVecT::TempStorage&>(smem_load_weight);
using vec_t = typename Ktraits::vec_t;
typename Ktraits::BlockLoadWeightVecT(smem_load_weight_vec).Load(
reinterpret_cast<vec_t*>(Bvar),
reinterpret_cast<vec_t(&)[Ktraits::kNLoads * 2]>(B_vals_load)
);
} else {
typename Ktraits::BlockLoadWeightT(smem_load_weight).Load(Bvar, B_vals_load, seqlen, 0.f);
}
#pragma unroll
for (int i = 0; i < kNItems; ++i) { B_vals[i] = complex_t(B_vals_load[i * 2], B_vals_load[i * 2 + 1]); }
}
}
template<typename Ktraits>
inline __device__ void store_output(typename Ktraits::input_t *out,
const float (&out_vals)[Ktraits::kNItems],
typename Ktraits::BlockStoreT::TempStorage &smem_store,
int seqlen) {
typename Ktraits::input_t write_vals[Ktraits::kNItems];
#pragma unroll
for (int i = 0; i < Ktraits::kNItems; ++i) { write_vals[i] = out_vals[i]; }
if constexpr (Ktraits::kIsEvenLen) {
auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_store);
using vec_t = typename Ktraits::vec_t;
typename Ktraits::BlockStoreVecT(smem_store_vec).Store(
reinterpret_cast<vec_t*>(out),
reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(write_vals)
);
} else {
typename Ktraits::BlockStoreT(smem_store).Store(out, write_vals, seqlen);
}
}
mamba/csrc/selective_scan/selective_scan_fwd_bf16.cu 0 → 100644
View file @ 2eefe3d6
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
// Split into multiple files to compile in paralell
#include "selective_scan_fwd_kernel.cuh"
template void selective_scan_fwd_cuda<at::BFloat16, float>(SSMParamsBase &params, cudaStream_t stream);
template void selective_scan_fwd_cuda<at::BFloat16, complex_t>(SSMParamsBase &params, cudaStream_t stream);
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment