* Improve the handling of quantized weights
Handling of quantized weights was split between two mechanisms:
- For quantized checkpoints, we used the new weight loader
infrastructure.
- For quantization while loading (EETQ, FP8, bitsandbytes) we
instead relied on conditional in `get_linear`.
Weight loaders support context managers to selectively load
particular layers with different weight loaders, which is useful
for models like Idefics2 AWQ, which uses a quantized text model,
but unquantized vision and connector models. However, the context
manager would be overrided by `get_linear`, which string-checks
`quantizer`. Also, the context manager would not work with
EETQ, FP8, and bitsandbytes.
This change migrates all quantizers to the weight loader infrastructure.
This has several benefits:
- We can use context managers with all quantizers.
- All the implementation details move down to the quantizer layers,
`get_linear` does not need to know how to handle quantizer linear
layers.
- All quantizer weights are strongly typed, we don't pass around
raw tensors.
- We don't have to pass around the `quantizer` string everywhere.
* Exclude non-MLP layers when using FP8 quantization with Llama
Use FP8 GPTQ-Marlin kernels to enable FP8 support on CUDA GPUs
with compute capability >=8.0 and <8.9.
Co-authored-by: Florian Zimmermeister <flozi00.fz@gmail.com>
GPTQ-Marlin is currently the best-performing kernel for GPTQ models. So
let's use it by default if the kernels are installed, the GPU supports
it, and the kernels support the configuration.
For models generated by `text-generation-server quantize`, use
`sym=False`. This subcommand symmetric quantization since the beginning
and incorrectly reporting the model to be symmetric will use
GPTQ-Marlin (which does not support asymmetric quantization).
This change adds support for 2:4 sparsity when using Marlin
quantization. The 2:4 kernel is used when:
* The quantizer is `marlin`;
* the quantizer checkpoint format is `marlin_24`.
Fixes#2098.
When the AWQ quantizer was used with a layer that uses a bias,
the bias tensor was not correctly passed/used. Instead, the
value `true`/`1.0` was added to the linear transformation.
Correctly pass through the bias when it is not `None`.
Fixes#2106.
Add support for GPTQ Marlin kernels
GPTQ Marlin extends the Marlin kernels to support common GPTQ
configurations:
- bits: 4 or 8
- groupsize: -1, 32, 64, or 128
- desc_act: true/false
Using the GPTQ Marlin kernels requires repacking the parameters in the
Marlin quantizer format.
The kernels were contributed by Neural Magic to VLLM. We vendor them
here for convenience.
This change adds support for Marlin-quantized models. Marlin is an
FP16xINT4 matmul kernel, which provides good speedups decoding batches
of 16-32 tokens. It supports quantized models with symmetric
quantization, groupsize -1 or 128, and 4-bit.
Tested with:
- Llama 2
- Llama 3
- Phi 3
# What does this PR do?
<!--
Congratulations! You've made it this far! You're not quite done yet
though.
Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.
Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.
Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->
<!-- Remove if not applicable -->
Fixes # (issue)
## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?
## Who can review?
Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.
<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @
@OlivierDehaene OR @Narsil
-->
Mostly straightforward, changes to existing code:
* Wrap quantizer parameters in a small wrapper to avoid passing
around untyped tuples and needing to repack them as a dict.
* Move scratch space computation to warmup, because we need the
maximum input sequence length to avoid allocating huge
scratch buffers that OOM.
Adds support for AMD Instinct MI300 in TGI.
Most changes are:
* Support PyTorch TunableOp to pick the GEMM/GEMV kernels for decoding
https://github.com/pytorch/pytorch/tree/main/aten/src/ATen/cuda/tunable.
TunableOp is disabled by default, and can be enabled with
`PYTORCH_TUNABLEOP_ENABLED=1`.
* Update ROCm dockerfile to PyTorch 2.3 (actually patched with changes
from https://github.com/pytorch/pytorch/pull/124362)
* Support SILU & Linear custom kernels contributed by AMD
* Update vLLM paged attention to https://github.com/fxmarty/rocm-vllm/,
branching out of a much more recent commit
3489ce7936
* Support FA2 Triton kernel as recommended by AMD. Can be used by
specifying `ROCM_USE_FLASH_ATTN_V2_TRITON=1`.
* Update dockerfile to ROCm 6.1
By default, TunableOp tuning results are saved in `/data` (e.g.
`/data/tunableop_meta-llama-Llama-2-70b-chat-hf_tp1_rank0.csv`) in order
to avoid to have to rerun the tuning at each `docker run`.
Example:
```
Validator,PT_VERSION,2.3.0
Validator,ROCM_VERSION,6.1.0.0-82-5fabb4c
Validator,HIPBLASLT_VERSION,0.7.0-1549b021
Validator,GCN_ARCH_NAME,gfx942:sramecc+:xnack-
Validator,ROCBLAS_VERSION,4.1.0-cefa4a9b-dirty
GemmTunableOp_Half_TN,tn_8192_7_28672,Gemm_Rocblas_45475,0.132098
GemmTunableOp_Half_TN,tn_10240_4_8192,Gemm_Rocblas_45546,0.0484431
GemmTunableOp_Half_TN,tn_32000_6_8192,Default,0.149546
GemmTunableOp_Half_TN,tn_32000_3_8192,Gemm_Rocblas_45520,0.147119
GemmTunableOp_Half_TN,tn_8192_3_28672,Gemm_Rocblas_45475,0.132645
GemmTunableOp_Half_TN,tn_10240_3_8192,Gemm_Rocblas_45546,0.0482971
GemmTunableOp_Half_TN,tn_57344_5_8192,Gemm_Rocblas_45520,0.255694
GemmTunableOp_Half_TN,tn_10240_7_8192,Gemm_Rocblas_45517,0.0482522
GemmTunableOp_Half_TN,tn_8192_3_8192,Gemm_Rocblas_45546,0.0444671
GemmTunableOp_Half_TN,tn_8192_5_8192,Gemm_Rocblas_45546,0.0445834
GemmTunableOp_Half_TN,tn_57344_7_8192,Gemm_Rocblas_45520,0.25622
GemmTunableOp_Half_TN,tn_8192_2_28672,Gemm_Rocblas_45475,0.132122
GemmTunableOp_Half_TN,tn_8192_4_8192,Gemm_Rocblas_45517,0.0453191
GemmTunableOp_Half_TN,tn_10240_5_8192,Gemm_Rocblas_45517,0.0482514
GemmTunableOp_Half_TN,tn_8192_5_28672,Gemm_Rocblas_45542,0.133914
GemmTunableOp_Half_TN,tn_8192_2_8192,Gemm_Rocblas_45517,0.0446516
GemmTunableOp_Half_TN,tn_8192_1_28672,Gemm_Hipblaslt_TN_10814,0.131953
GemmTunableOp_Half_TN,tn_10240_2_8192,Gemm_Rocblas_45546,0.0481043
GemmTunableOp_Half_TN,tn_32000_4_8192,Gemm_Rocblas_45520,0.147497
GemmTunableOp_Half_TN,tn_8192_6_28672,Gemm_Rocblas_45529,0.134895
GemmTunableOp_Half_TN,tn_57344_2_8192,Gemm_Rocblas_45520,0.254716
GemmTunableOp_Half_TN,tn_57344_4_8192,Gemm_Rocblas_45520,0.255731
GemmTunableOp_Half_TN,tn_10240_6_8192,Gemm_Rocblas_45517,0.0484816
GemmTunableOp_Half_TN,tn_57344_3_8192,Gemm_Rocblas_45520,0.254701
GemmTunableOp_Half_TN,tn_8192_4_28672,Gemm_Rocblas_45475,0.132159
GemmTunableOp_Half_TN,tn_32000_2_8192,Default,0.147524
GemmTunableOp_Half_TN,tn_32000_5_8192,Default,0.147074
GemmTunableOp_Half_TN,tn_8192_6_8192,Gemm_Rocblas_45546,0.0454045
GemmTunableOp_Half_TN,tn_57344_6_8192,Gemm_Rocblas_45520,0.255582
GemmTunableOp_Half_TN,tn_32000_7_8192,Default,0.146705
GemmTunableOp_Half_TN,tn_8192_7_8192,Gemm_Rocblas_45546,0.0445489
```
---------
Co-authored-by: Mohit Sharma <mohit21sharma.ms@gmail.com>
This PR adds paligemma modeling code
Blog post: https://huggingface.co/blog/paligemma
Transformers PR: https://github.com/huggingface/transformers/pull/30814
install the latest changes and run with
```bash
# get the weights
# text-generation-server download-weights gv-hf/PaliGemma-base-224px-hf
# run TGI
text-generation-launcher --model-id gv-hf/PaliGemma-base-224px-hf
```
basic example sending various requests
```python
from huggingface_hub import InferenceClient
client = InferenceClient("http://127.0.0.1:3000")
images = [
"https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png",
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png",
]
prompts = [
"What animal is in this image?",
"Name three colors in this image.",
"What are 10 colors in this image?",
"Where is the cow standing?",
"answer en Where is the cow standing?",
"Is there a bird in the image?",
"Is ther a cow in the image?",
"Is there a rabbit in the image?",
"how many birds are in the image?",
"how many rabbits are in the image?",
]
for img in images:
print(f"\nImage: {img.split('/')[-1]}")
for prompt in prompts:
inputs = f"{prompt}\n"
json_data = {
"inputs": inputs,
"parameters": {
"max_new_tokens": 30,
"do_sample": False,
},
}
generated_output = client.text_generation(prompt, max_new_tokens=30, stream=False)
print([f"{prompt}\n{generated_output}"])
```
---------
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
# What does this PR do?
<!--
Congratulations! You've made it this far! You're not quite done yet
though.
Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.
Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.
Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->
<!-- Remove if not applicable -->
Fixes # (issue)
## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?
## Who can review?
Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.
<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @
@OlivierDehaene OR @Narsil
-->
Mohit Sharma
