[Dynamo] Small enchancements for graph dump ir and task arguments#172
[Dynamo] Small enchancements for graph dump ir and task arguments#172yaoyaoding merged 33 commits intohidet-org:mainfrom
Conversation
xinli-git
commented
Apr 13, 2023
- When graphs are partitioned by dynamo, the previous graph gets overwritten. Using a "function static" variable to track graph sequence
- Temporary support for passing along additional parameters to the function, this change is used for better support for third party implementations where shape info or dynamic attributes may be passed in.
This reverts commit a1e8df0.
yaoyaoding
left a comment
There was a problem hiding this comment.
Thanks @xinli-git. I left some feedbacks.
| generate_executor.graph_counter = getattr(generate_executor, 'graph_counter', 0) + 1 | ||
| save_dir: Path = Path(save_dir) | ||
| ctx.save_graph_instrument(save_dir / "graph_{}".format(generate_executor.graph_counter)) |
There was a problem hiding this comment.
How do you think to implement this as a seperate function?
Besides, I prefer to find the first non-existed directory {save_dir}/<id> as the save_dir instead of track the id internally.
There was a problem hiding this comment.
the thing with find the first non-existed directory is that if I recompile the model and use the same save_dir value, it will not overwrite the previous compiled graphs, but keeps incrementing. I didn't like this behaviour.
If you think it is OK to disallow overwriting, then I am happy to make it find the first non-existed directory
How do you think to implement this as a seperate function?
- you mean make this three lines into a different function?
There was a problem hiding this comment.
the thing with find the first non-existed directory is that if I recompile the model and use the same save_dir value, it will not overwrite the previous compiled graphs, but keeps incrementing. I didn't like this behaviour.
I see. Make sense.
How do you think to implement this as a seperate function?
Yes. For example, something like
def _resolve_save_dir(save_dir: str) -> str:
_resolve_save_dir.counter = _resolve_save_dir.counter = getattr(..., 'counter', 0) + 1
return str(Path(save_dir) / "graph_{}".format(_resolve_save_dir.counter))
python/hidet/graph/operator.py
Outdated
| """An operator that takes tensor as input and output.""" | ||
|
|
||
| def __init__(self, inputs: List[Tensor], attributes: Dict[str, Any], task: Optional[Task]): | ||
| def __init__(self, inputs: List[Tensor], attributes: Optional[Dict[str, Any]] = None, task: Optional[Task] = None): |
There was a problem hiding this comment.
I prefer to make attributes non-optional. If it is empty, we force the developer to give an empty dict {} in case they forget to pass attributes (if they forget to do so, the reforward function would fail).
For the task attribute, let's also use the Optional[Task] but do not give the default value. In case of the 3rdparty kernel, we explicitly pass the None to task.
There was a problem hiding this comment.
OK that sounds good, I added this change because the code below checks if it is None and assigns it to an empty dict if that's the case. This gave me the wrong impression that you meant to have this optional but forgot to do so
There was a problem hiding this comment.
I see. In the very beggining, I set the default value of attributes=None, and some operators that should pass the attributes did not pass them. So I make this parameter mandatory later, but forget to change the body of init. Feel free to delete the check of attributes==None.
|
It seems that the github action vm is acting weird again |
Our the nvidia driver on our ci server has done. Fixed just now. |
|
Thanks @xinli-git ! |
…det-org#172) * exp, float * wip * chunk, groupnorm, softmax, baddbmm, emmpty * add interpolate, lint and format * revert changes of import hidet at top level to minimize changes for PR * typo * trigger actions * trigger actions * dummy commit * dummy commit * add some optimizations to skip certain operations based on alpha beta * add group norm test * format * introduce a fix to torch.compile not dumping graph IR * Revert "introduce a fix to torch.compile not dumping graph IR" This reverts commit a1e8df0. * add interlolate test and group norm test * accidental push * remove a random newline added * minor hot fix * add a function level static var for multi-graph save ir * small line change * revert previous changes * move save dir to new utility function * format lint and remove optional attr * trigger actions * trigger actions * Update operator.py --------- Co-authored-by: Xin Li <xin@centml.ai> Co-authored-by: Yaoyao Ding <dingyaoyao.cs@gmail.com>