[Neuron] Adding support for adding/ overriding neuron configuration a… #8062

hbikki · 2024-09-01T11:05:57Z

…nd adding support for neuron model quantization configuration.

Quantization : Added support for int8 quantization for the neuron models, this is achieved by passing the quant config as a param tot he neuron config, the NEURON_QUANT_DTYPE helps us customize the dtype for the quantization.
Neuron Config: Update the default neuron config to support the latest neuron release's optimizations out of the box.
Dynamic Config update/Override : Provided support to override or pass neuron config from the engine args -> ModelConfig -> NeuronConfig , this will minimize the updates to vllm and provides more customizable configuration for clients.

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

[Bugfix] for bug fixes.
[CI/Build] for build or continuous integration improvements.
[Doc] for documentation fixes and improvements.
[Model] for adding a new model or improving an existing model. Model name should appear in the title.
[Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
[Kernel] for changes affecting CUDA kernels or other compute kernels.
[Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
[Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
[Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

We adhere to Google Python style guide and Google C++ style guide.
Pass all linter checks. Please use format.sh to format your code.
The code need to be well-documented to ensure future contributors can easily understand the code.
Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

github-actions · 2024-09-01T11:06:10Z

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

Comment /ready on the PR
Add ready label to the PR
Enable auto-merge.

🚀

hbikki · 2024-09-02T01:28:15Z

/ready

liangfu

Thanks @hbikki for contribuing. The proposed change looks good overall. I left some comments.

liangfu · 2024-09-03T17:35:31Z

vllm/model_executor/model_loader/neuron.py

+        "quant":
+        neuron_quantization_config_builder(model_config.quantization)
+        if model_config.quantization else None,


we will need weight_tiling flag, when quantization being enabled.

yep added weight_tiling, thanks!

liangfu · 2024-09-03T17:40:35Z

vllm/model_executor/layers/quantization/neuron_quant.py

+        dequant_dtype: str = "f16",
+        quantize_method: str = "vector_dynamic",
+    ) -> None:
+        self.quant_dtype = os.getenv("NEURON_QUANT_DTYPE", "s8")


is this configurable from existing vLLM config? can we pass existing vLLM config ?

the datatype we use for quantization in neuron doesn't match the torch d types or existing quant , so created a new quant config to support it.

if we could map torch dtype to neuron dtype, we should be able to avoid inventing the environment variable specifically for neuron backend.
for instance, s8 can be translated to torch.int8, and f8e4m3fn and be translated to torch.float8_e4m3fn.
ideally, we should be able to configure quantization data type with existing vllm config.

help map the relation between neuron dtype and torch dtype?

liangfu · 2024-09-03T17:41:46Z

vllm/model_executor/layers/quantization/neuron_quant.py

+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+    def get_tnx_quantization_config(self):


avoid tnx terminology? it will be confusing what is tnx.

got it , as the quant config is specific to tnx, I can do a tnx package check before using the config , any thoughts ?

liangfu · 2024-09-03T17:55:54Z

vllm/model_executor/model_loader/neuron.py

+    neuron_quantization_config_builder = lambda quant: get_quantization_config(
+        quant).from_config(quant_config).get_quant_method(None, "")
+    # TODO: Add Paged attention config to the default neuron arguments.
+    default_neuron_args = {


1/ Can we use layout enum instead of strings (e.g. BSH)?
2/ Can we build dict with assignment statements?

This would be something like the following:

default_neuron_config_dict = dict( fused_qkv=True, attention_layout=constants.Layout.BSH, ...)

updated for better readability, thanks!

liangfu · 2024-09-03T17:58:39Z

vllm/model_executor/model_loader/neuron.py

+                                      overridden_neuron_config):
+    from transformers_neuronx.config import NeuronConfig
+    overridden_neuron_config = overridden_neuron_config or {}
+    combined_config = {**default_neuron_config, **overridden_neuron_config}


what happens if there is an overlap?

The overridden config replaces the default neuron config

example : when use use this as override config:

override_neuron_config={"cast_logits_dtype":"bfloat16", "collectives_layout":"Dummy"}

we get this result :

(Pdb) combined_config {'collectives_layout': 'Dummy', 'attention_layout': 'BSH', 'fuse_qkv': True, 'quant': <transformers_neuronx.config.QuantizationConfig object at 0x7f259835a2f0>, 'continuous_batching': <transformers_neuronx.config.ContinuousBatchingConfig object at 0x7f259831d750>, 'weight_tiling': True, 'cast_logits_dtype': 'bfloat16'}

for this line

combined_config = {**default_neuron_config, **overridden_neuron_config}

there would be an error when we override like this:

combined_config={"collectives_layout":"BSH", "collectives_layout":"HSB"}

right?

no the dict supports only one key with the name , the second collectives_layout will overwrite the first one ,
Based on our offline sync update to .update() for readability.

liangfu · 2024-09-04T19:35:05Z

vllm/model_executor/model_loader/neuron.py

+        quant=neuron_quantization_config_builder(model_config.quantization)
+        if model_config.quantization else None,
+        continuous_batching=continuous_batching_config,
+        weight_tiling=True)


we might not want weight_tiling enabled, when weight quantization is not enabled.
pair weight_tiling with weight quantization config ?

Sure, I couldn't find an env variable to enable it dynamically , so paired it with qunat

liangfu · 2024-09-04T19:46:18Z

vllm/model_executor/model_loader/neuron.py

@@ -119,19 +136,51 @@ def _get_buckets(env: str, default_value: List[int]) -> List[int]:
    return buckets_list


+def _get_default_tnx_neuron_config(model_config: ModelConfig,


also avoid tnx terminology here?

…nd adding support for neuron model quantization configuration.

liangfu

Thanks @hbikki for the update. It looks good.

I think we can leave the dtype mapping between neuron-specific and torch native data types as a follow-up.

vllm-project#8062) Co-authored-by: Harsha Bikki <[email protected]> Signed-off-by: Alvant <[email protected]>

vllm-project#8062) Co-authored-by: Harsha Bikki <[email protected]> Signed-off-by: LeiWang1999 <[email protected]>

hbikki force-pushed the neuron-config branch 4 times, most recently from e301a9f to f1688eb Compare September 2, 2024 00:24

github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label Sep 2, 2024

WoosukKwon added the aws-neuron Related to AWS Inferentia & Trainium label Sep 2, 2024

liangfu reviewed Sep 3, 2024

View reviewed changes

hbikki force-pushed the neuron-config branch 2 times, most recently from 4b6a00b to f4ac863 Compare September 4, 2024 17:46

liangfu reviewed Sep 4, 2024

View reviewed changes

hbikki force-pushed the neuron-config branch from f4ac863 to 1e256ad Compare September 4, 2024 19:41

liangfu reviewed Sep 4, 2024

View reviewed changes

hbikki force-pushed the neuron-config branch from 1e256ad to 10da4f3 Compare September 4, 2024 19:58

[Neuron] Adding support for adding/ overriding neuron configuration a…

fbae9ad

…nd adding support for neuron model quantization configuration.

hbikki force-pushed the neuron-config branch from 10da4f3 to fbae9ad Compare September 4, 2024 20:53

hbikki requested a review from liangfu September 4, 2024 21:51

liangfu approved these changes Sep 4, 2024

View reviewed changes

simon-mo approved these changes Sep 4, 2024

View reviewed changes

simon-mo merged commit 008cf88 into vllm-project:main Sep 4, 2024
53 checks passed

Alvant pushed a commit to compressa-ai/vllm that referenced this pull request Oct 26, 2024

[Neuron] Adding support for adding/ overriding neuron configuration a… (

619a49c

vllm-project#8062) Co-authored-by: Harsha Bikki <[email protected]> Signed-off-by: Alvant <[email protected]>

AlpinDale mentioned this pull request Dec 24, 2024

neuron: add 8bit quantization for Neuron aphrodite-engine/aphrodite-engine#994

Merged

		@@ -119,19 +136,51 @@ def _get_buckets(env: str, default_value: List[int]) -> List[int]:
		return buckets_list


		def _get_default_tnx_neuron_config(model_config: ModelConfig,

Uh oh!

[Neuron] Adding support for adding/ overriding neuron configuration a… #8062

[Neuron] Adding support for adding/ overriding neuron configuration a… #8062

Uh oh!

Conversation

hbikki commented Sep 1, 2024

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

Uh oh!

github-actions bot commented Sep 1, 2024

Uh oh!

hbikki commented Sep 2, 2024

Uh oh!

liangfu left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liangfu Sep 4, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liangfu Sep 4, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liangfu left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

liangfu Sep 4, 2024 •

edited

Loading

liangfu Sep 4, 2024 •

edited

Loading