Upstream merge Nov20 (includes ft_group quantization support)
#71
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Fix two bugs in kv-cache pop loop Bug 1: old code would stop early because output_ids was shortened in-place during the loop Bug 2: off-by-one in backoff size due to break
…1017) This commit adds an optional `--pdb` flag to the `build.py` script. If passed, any exception raised that would otherwise terminate the script will first enter a pdb post-mortem, allowing the error to be inspected.
…ai#1040) Add doc for ChatConfig, ConvConfig, GenerationConfig, BuildArgs, build model
llama2 q4f160
fix permission issue
Support for the stablelm-3b-4e1t model
* Iterate model prebuilts docs * small fix
This PR separates out the tokenizer creation function, the random number generator out from `llm_chat.cc` as a preparation step for batching inference support, since these functions/modules are also used in the same way in batching inference.
Update README.md
* add verbose stats to mlc-chat REST API * update docs
* [Transform] Apply split_rotary optimization on prefill Prior to this commit, the `transform.fuse_split_rotary_embedding` function was only applicable to the `decode` function of a Llama-type model. This was due to the sequence length being restricted to one, both in the pattern-match rule and in the `split_rotary` function, and the function being restricted to operate only on the `decode` function. This commit updates the `transform.fuse_split_rotary_embedding` pass to be a `tvm.ir.transform.Pass`, operating on all applicable matched in the `IRModule`. The `split_rotary` function is now produced as a fully-generic function, with static parameters substituted in afterwards. At this stage, the sequence length is retained as a dynamic parameter, such that it can be used by the `prefill` function. * Avoid multiple kernel launches for split_rotary
…i#1055) Co-authored-by: Junru Shao <junrushao1994@gmail.com>
…i#1033)" (mlc-ai#1058) This reverts commit b9179cf as elaborated here mlc-ai#1033 (comment)
…ma-2 families (mlc-ai#1032) * fix * reflect feedback --------- Co-authored-by: “Sunghyun <sunggg@umich.com>
`--force-reinstall` will reinstall all dependencies to a python package, which is unnecessary. `-U` is a better choice in this case.
This PR introduces the initial batched input support for llama models. To make the code managable, we keep both the single-sequence handling flow and the batching handling flow in the Llama modeling. Now, with `--enable-batching` as a build argument, we build Llama for the batched version. NOTE: The paged attention kernel/TIR func are not included in this PR, so currently the built library with batching enabled is not runnable. We will follow up with the attention kernel in the future. This PR guarantees that the existing single-sequence inference (Python API, CLI, etc.) is not broken. P.S.. The batching flow is subject to bug fixes as we integrate with the attention function and run the e2e flow in the future.
* [stablelm 3b] Rename dynamic vocab size from "v" to "vocab_size" * Add get_num_key_value_heads method to StableLM3bConfig
This commit removes the `if`/`elif` chain in `core.py`, where the body of each conditional assigns the same `mod, param_manager, params, model_config`, and is identical except for the choice of model being built.
This commit replaces the single-parameter `relax_model.param_manager.create_quantize_func` function with a method on the `ParamManager`, `create_parameter_transformation`. This avoids potential typos between `param_manager` as the imported Python module `mlc_llm.relax_model.param_manager` and an instance of the `ParamManager` class named `param_manager`, and makes the functionality easier to find. This function also takes an optional `optimize_parameter_order` flag, defaulting to `True`, which applies the `ReorderTransformFunc` pass. Since the `ReorderTransformFunc` is intended to be used with several configuration objects owned by `ParamManager`, this simplifies the common path of producing an optimally-ordered parameter transformation module.
PR mlc-ai#1048 updated the signature of softmax in the built model library and changed the temperature buffer shape in ChatModule. This causes some existing demo unable to run since we did not do a round of model library update. This PR reverts the ChatModule change, and adds back the softmax function in non-batching case. With this PR, the regression should be fixed.
…ai#1074) This PR lifts the device string parsing (just a few of lines) to a standalone function, so that on the serving side the serving can make use of this function as well. Tested Python API and it does not seem to incur regression.
The pass `fuse-split-rotary` assumes the compute dtype is fp16, which usually is, but in certain cases, e.g. `q0f32` and `q4f32_1`, the compute is based on fp32 instead. This PR strengthens the check guard.
This PR establishes the compiler components in MLC-Chat Python API, which currently includes two primary components: models and parameters. The models are `nn.Module`-based definition of an LLM, which, as the very first stab, contains only `LlamaForCasualLM`. It is decomposed into three files: - `llama_config.py`: common configurations for Llama, where we define relevant configurations of its architecture, as well as include standard config file for Llama2-7B/13B/70B for convenient testing; - `llama.py`: the model architecture of Llama, based on the PyTorch-like `nn.Module` API; - `llama_parameter.py`: defines the mapping between MLC parameters and pytorch parameters. The parameters contains the basic functionality of parameter mapping, and the loaders that effectively convert parameters from PyTorch to MLC according to the mapping specified. Currently, only `HFTorchLoader` is implemented, but loaders like SafeTensor, GPTQ or AWQ should be quite straightforward according to the existing design. On top of this PR, on-the-fly quantization could be defined as a loading time transformation on MLC parameters, while pre-quantized parameter loading is effectively parameter loading after MLC's `nn.Module` is quantized. Two unittests examplify how the infrastructure works: - `./tests/python/model/test_llama.py` shows how to create an `nn.Module` using the new infra, and then convert it to TVM IRModule; - `./tests/python/parameter/hf_torch_loader.py` shows how to load parameters from HuggingFace PyTorch format. Besides, `mlc_chat.support` is established for utility functions, which now contains two utils: - `config.py` which supports reading configurations into dataclasses from JSON file or Python dict. On top of Python dataclass, it throws irrelevant fields into `cls.kwargs`, which is helpful when loading HuggingFace configuration file; - `tqdm.py` which contains tqdm-related utilities, primarily redirecting logging and printing to work nicely with tqdm.
This PR makes it possible to invoke mlc_chat subcommands directly. Previously one has to use `python -m` as the prefix to invoke `mlc_chat`: ```bash python -m mlc_chat compile \ --model /models/Llama-2-7b-chat-hf \ --quantization q4f16_1 \ --max-sequence-length 4096 \ -o ./llama.so ``` This PR makes is possible to use it without the `python -m` prefix: ```bash mlc_chat compile \ --model /models/Llama-2-7b-chat-hf \ --quantization q4f16_1 \ --max-sequence-length 4096 \ -o ./llama.so ```
…-ai#1262) Following PR mlc-ai#1253, I think ergonomics of the `generate` function of `ChatModule` can be improved (given it's an important public-facing API). This PR simplifies the function's usage by implementing the `From` trait for the `Prompt` enum. Also updated the example code. Now the interface changes to: ```rust /// Single prompt case: cm.generate("what is the meaning of life?", None) /// Multiple prompt case: let messages: Vec<ChatMessage> = vec![message1, message2, message3]; let output = cm.generate(messages, None).unwrap(); ```
"Context window" is a terminology better aligned with LLM world. Whenever a new model is trained, it is one of the most important metrics that people care about. Therefore, I'd love to switch it over sooner than later, before "mlc_chat compile" becomes mature and documented.
This PR fixes the shape infer for group quantization.
This PR provides a script that automatically quantizes models from
HuggingFace using various quantization formats as specified.
Example: When being provided the following JSON file:
```json
{
"destination": "{username}/{model_id}-{quantization}", # Name of HF repo
"default_quantization": ["q0f16", "q0f32", "q3f16_1", "q4f16_1", "q4f32_1"],
"tasks": [
{
"model_id": "Llama-2-7b-hf",
"model": "/models/Llama-2-7b-hf", # Can be HF URL or a local path
"context_window_size": 4096,
"conv_template": "LM",
"quantization": [
{
"format": "q4f16_awq",
"model": "https://huggingface.co/TheBloke/Llama-2-7B-AWQ", # Overriding default `source`
"source_format": "awq"
}
]
}
]
}
```
The script will automatically run quantization and upload them to the
following repos:
- https://huggingface.co/junrushao/Llama-2-7b-hf-q0f16
- https://huggingface.co/junrushao/Llama-2-7b-hf-q0f32
- https://huggingface.co/junrushao/Llama-2-7b-hf-q3f16_1
- https://huggingface.co/junrushao/Llama-2-7b-hf-q4f16_1
- https://huggingface.co/junrushao/Llama-2-7b-hf-q4f32_1
- https://huggingface.co/junrushao/Llama-2-7b-hf-q4f16_awq
This PR introduces a few enhancements: - Allow to override temporary path via environment variable `MLC_TEMP_DIR`; - Add a 10-time retry when uploading the quantized weights to HuggingFace Hub. It could fail at times; - Echo the commands being used to quantize the models in `logs.txt`; - Fix a compatibility issue when pulling individual weights down from HuggingFace Hub in Git LFS.
* add python, rest api test * remove mistral, fix pylint * fix pylint requests import error
This fix enables default models in app-config.json to get shown "downloaded" in model list via with adb push method for the default models
This PR unifies automatic device detection logic by using `mlc_chat.support.auto_device`, which comes with detailed logging and fallback mechanisms.
This PR fixes the broken CI due to different tasks sharing the same workspace.
* generalize `prefill-chunk-size` * renaming `cache_len` to `rolling_cache_len` * [nn.Module] generalize `prefill_chunk_size` * quick fix * lint fix * check sw with chunking * fix `_attach_variable_bounds` * update config from lib metadata * cleanup cleanup * metadata fix
compatible for chatglm
* Add q4/q8_ft_group quantization mode * Update submodule
masahi
changed the title
ft_group quantization support)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
To use mlc-ai#1284
The accompanying TVM hash: 605a9ad8ce8b81710e626ad7ebf8a8167fbbfaff from https://github.com/masahi/tvm/tree/for-mlc-serve-nov22:
@sunggg @binarybana