Trainer: the model analysis on the AOT compiled JAX program. #1036
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
|
@markblee Could you review it? From 1112 |
apghml
requested changes
Mar 5, 2025
| if not hasattr(compiled, "memory_analysis"): | ||
| return "" | ||
|
|
||
| to_mb_gb = lambda x: f"{x / (1024**2):.1f} MB / {x / (1024**3):.2f} GB" |
There was a problem hiding this comment.
I think this could be confusing since users might think the left number is the usage and right number is the maximum available. (e.g., they might mistake it for meaning "x mb out of y gb used"). Can we eliminate the mb?
There was a problem hiding this comment.
But some downstream model (e.g. speech model), GB is too big. Let me make it dynamically decide MB or GB.
| @@ -91,6 +101,14 @@ def _compile_and_dump_programs( | |||
| logging.info("Wrote serialized %s to %s", program_name, serialized_compiled_output_path) | |||
There was a problem hiding this comment.
This PR still doesn't deduplicate the the memory printing code in run_aot_compilation.py with the new function you have in trainer.py?
This will help researchers estimate HBM usage and computation costs before launching a job, allowing them to determine whether a model is compute-bound or memory-bound. Introduced aot_model_analysis(), which returns analysis results as a string, making it reusable (e.g., in Jupyter notebooks). `run_aot_compilation` tool prints fuji-1B-v3 model analysis as follows. ``` ======= Memory Analysis ================================== Input memory: 4465.0 MB / 4.36 GB Output memory: 4464.8 MB / 4.36 GB Temp memory: 174977.1 MB / 170.88 GB Code memory: 0.0 MB / 0.00 GB Total HBM memory: 183906.9 MB / 179.60 GB ======= Cost Analysis ==================================== FLOPS: 71733280.0 M / 70052.03 G The number of exp/log/sin/cos ops: 21364.8 M / 20.86 G The total memory traffic: 1792723.2 MB / 1750.71 GB HBM access: 751479.1 MB / 733.87 GB L2 cache access: 328740.8 MB / 321.04 GB Register usage: 61266.7 MB / 59.83 GB Output data transferred: 677251.9 MB / 661.38 GB Hardware utilization scores Tensor Cores / MatMul units: 647.0 ALU (Arithmetic Logic Unit): 430.0 Memory Load/Store Units: 144.0 L1 Cache Operations: 92.0 L2 Cache Operations: 60.0 Special Function Units (exp/log/sin/cos): 41.0 Integer Units (for indexing, loop counters): 16.0 Branch Divergence (Control Flow Processing): 12.0 Load Balancing / Dispatch): 10.0 Texture Units (or Rarely Used Compute Units): 8.0 ```
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.
This will help researchers estimate HBM usage and computation costs before launching a job, allowing them to determine whether a model is compute-bound or memory-bound.
Introduced aot_model_analysis(), which returns analysis results as a string, making it reusable (e.g., in Jupyter notebooks).
In addition, change
run_aot_compilationto support it on CPU.run_aot_compilationtool prints fuji-1B-v3 model analysis as follows.