Parallelise Model Loading #466
Open
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
|
@AlexCheema PTAL lmk if u need more changes. |
AlexCheema
requested changes
Nov 18, 2024
|
@AlexCheema PTAL |
AlexCheema
requested changes
Nov 20, 2024
- Wrapped debug print statements with `DEBUG >= 2` condition for better logging control. - Consolidated shard building and preloading into a single operation using list comprehensions. - Improved error handling to cover all models in a batch, reducing redundancy. - Added clearer messaging for unsupported models. - Simplified code structure for better readability and performance.
|
@AlexCheema , say do I run the formatter over the whole codebase ?? or just the files I edited ? |
AlexCheema
requested changes
Nov 21, 2024
| @@ -1,4 +1,4 @@ | |||
| from typing import Optional, Tuple, TYPE_CHECKING | |||
| rom typing import Optional, Tuple, TYPE_CHECKING | |||
| @@ -133,17 +134,21 @@ | |||
| lambda req_id, tokens, __: topology_viz.update_prompt_output(req_id, inference_engine.tokenizer.decode(tokens)) if topology_viz and hasattr(inference_engine, "tokenizer") else None | |||
| ) | |||
|
|
|||
| def preemptively_start_download(request_id: str, opaque_status: str): | |||
|
|
|||
| async def preemptively_start_download(request_id: str, opaque_status: str): | |||
There was a problem hiding this comment.
I don't think this should be made async.
creating tasks fire-and-forget-style like it was before is better. if you need to run stuff in sequence, create a task for an asynchronous function that awaits each step.
| asyncio.create_task(shard_downloader.ensure_shard(current_shard, inference_engine.__class__.__name__)) | ||
| await shard_downloader.ensure_shard(current_shard, inference_engine.__class__.__name__) | ||
| await node.preload_models([current_shard]) | ||
| return current_shard |
Comment on lines
+240
to
+243
| shards = [ | ||
| shard for model in models_to_preload | ||
| if (shard := build_base_shard(model, inference_class)) is not None | ||
| ] |
There was a problem hiding this comment.
but these on one line please
Comment on lines
+246
to
+249
| unsupported = [ | ||
| model for model in models_to_preload | ||
| if not build_base_shard(model, inference_class) | ||
| ] |
Comment on lines
+236
to
+237
| if DEBUG >= 2: | ||
| print(f"Preloading models: {models_to_preload}") |
|
Please respond to my review @vovw |
flooded with college work rn will address these tomorrow |
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.
Add model parallelise preloading capability for improved inference startup time
Adds model preloading functionality to improve initial inference latency by allowing models to be loaded into memory before they're needed.
Key changes:
--preload-modelsCLI arg to specify models for preloadingpreload_modelmethod in inference engine interfacePrimary motivation is reducing cold-start latency by preloading models before they're needed, useful for deployments requiring predictable latency.
Tested with MLX engine and verified working with preemptive downloads. Built on existing shard infrastructure, maintains backward compatibility.
test using
prev pr #360