Draft
Conversation
Small fix for querying service
Add example for serving Llama 3 8B
Use L4 instead of A10G for LLM serving example
Can't actually use `pathlib.Path` because that's only designed for file paths not URLs :/ But `urllib.parse.urljoin` works and is a bit prettier (it's a standard python library) --------- Signed-off-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Aydin Abiar <aydin@anyscale.com>
# Why this change We previously pinned to vLLM v0 due to a bug in Ray 2.48.0 that blocked Hugging Face tokens from runtime dependencies. This is bad because non-maintanable (vLLM v0 will be deprecated soon). I just found out we can pass the token directly through engine parameters, so we can now use vLLM v1 with gated models. # Summary * Switched deployment from vLLM v0 → vLLM v1. * Added a C compiler to the minimal Dockerfile since vLLM v1 depends on Triton, which compiles C code at runtime ( see vllm-project/vllm#2997 ). * Resolved Hugging Face token issue by passing it directly via engine parameters instead of runtime dependencies. * Updated model to [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) (more popular for most usage). --- # Testing * Tested Anyscale Service on AWS cloud. --------- Signed-off-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Robert Nishihara <robertnishihara@gmail.com> Co-authored-by: Robert Nishihara <rkn@anyscale.com>
Multiple changes here... **README** Updated per Douglas’s review ([https://github.com/anyscale/docs/pull/1464](https://github.com/anyscale/docs/pull/1464)): * Added a description at the top. * Applied code highlighting to bash commands. * Made `HF_TOKEN` usage clearer (explicit `export` example) and added instructions for ungated models. * Clarified where to set token/endpoint before querying the service. * Added `pip install openai` requirement before querying * Rephrased future tense and passive voice. **serve\_llama\_3\_1\_8b.py** With Ray 2.49.0 we can now forward the Hugging Face token to vLLM v1 via runtime dependencies: * Pass `HF_TOKEN` as a runtime dependency instead of engine parameters. * Updated ungated model suggestion to use Unsloth’s Llama variant (so we stay within the Llama family instead of switching to Qwen). **Dockerfile** Simplified for Ray 2.49.0: * No need to pin `transformers==4.53.3` anymore * Removed `uv` installation; if the goal is to build a minimal image based on ray then let's remove uv ? This might confuse users --------- Signed-off-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Robert Nishihara <rkn@anyscale.com>
Add a tutorial on deploying llama 3.1 8b-instruct. Follow the same format as the llama 3.1 8b example for consistency (see this PR: [https://github.com/anyscale/examples/pull/12](https://github.com/anyscale/examples/pull/12)). As with the reasoning for choosing L4 GPUs in the Llama 3 8B example, here we use 8×A100 GPUs, which are available in both our AWS and GCP clouds. --------- Signed-off-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Aydin Abiar <aydin@anyscale.com> Co-authored-by: Robert Nishihara <rkn@anyscale.com>
…ples into nemo-curator-dedup
| return asyncio.run(process_batch(batch, output_dir, batch_num)) | ||
|
|
||
|
|
||
| def download_webdataset( |
Contributor
There was a problem hiding this comment.
This assumes the whole dataset fits on disk in one machine, right? (Fine for LAION since it is just URLs, but probably not in general.)
What's the best way to get data into NeMo Curator? E.g., would it make sense to use Ray Data to read the data and stream it in? Or does NeMo Curator have methods for this?
Contributor
Author
There was a problem hiding this comment.
Since nemo curator uses nvidia DALI, I think the ideal data loading story would be to have all the images in something like s3 partitioned into different tar shards. We can then mount the s3 on each of the nodes, with each node accessing the subset of tar shards that it is computing on. Would you like me to build this into the example?
IanDJordan
force-pushed
the
nemo-curator-laion
branch
from
aa759ac to
d209f5f
Compare
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
No description provided.