[jit] Fix value association with dictionaries in the tracer #40885
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
`TracingState::setValue` associates a concrete IValue in the traced program with a `Value*` symbolic. Previously, the logic for how GenericDicts worked was special cased to only work for very simple cases and silently eat other cases. This PR generalizes the logic to reflect the same behavior as using dictionaries on input: whenever we encounter a dictionary in the system, we completely "burn in" all the keys into the graph, and then recursively call `setValue` on the associated value. This has the effect of requiring that any dictionary structure you are creating in a traced program be of fixed structure, similar to how any dictionary used as input must be static as well. [ghstack-poisoned]
facebook-github-bot
added
the
oncall: jit
💊 CI failures summary and remediationsAs of commit 4aba185 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 3 times. |
wanchaol
approved these changes
Jul 1, 2020
| } | ||
| if (value_type->cast<TensorType>()) { | ||
| handle_unpack(aten::values); | ||
| for (const auto& entry : dict) { |
There was a problem hiding this comment.
This will actually implies that if dict[tensor, tensor] (tensor as keys) as input or cross tracing/scripting, the tensor key will always be treated as constant, which is ok, but I am wondering if we should throw some warning here to user?
There was a problem hiding this comment.
Agreed, I filed #40886 as a followup to improve the error messaging in all cases where a key mismatch occurs
| auto static_key = graph->insertConstant(entry.key()); | ||
| auto static_value = | ||
| graph->insert(aten::__getitem__, {value, static_key}); | ||
| setValue(entry.value(), static_value); | ||
| } | ||
| } else { | ||
| std::ostringstream os; |
There was a problem hiding this comment.
error message seems out dated
`TracingState::setValue` associates a concrete IValue in the traced program with a `Value*` symbolic. Previously, the logic for how GenericDicts worked was special cased to only work for very simple cases and silently eat other cases. This PR generalizes the logic to reflect the same behavior as using dictionaries on input: whenever we encounter a dictionary in the system, we completely "burn in" all the keys into the graph, and then recursively call `setValue` on the associated value. This has the effect of requiring that any dictionary structure you are creating in a traced program be of fixed structure, similar to how any dictionary used as input must be static as well. Differential Revision: [D22342490](https://our.internmc.facebook.com/intern/diff/D22342490) [ghstack-poisoned]
suo
added a commit
that referenced
this pull request
Jul 1, 2020
`TracingState::setValue` associates a concrete IValue in the traced program with a `Value*` symbolic. Previously, the logic for how GenericDicts worked was special cased to only work for very simple cases and silently eat other cases. This PR generalizes the logic to reflect the same behavior as using dictionaries on input: whenever we encounter a dictionary in the system, we completely "burn in" all the keys into the graph, and then recursively call `setValue` on the associated value. This has the effect of requiring that any dictionary structure you are creating in a traced program be of fixed structure, similar to how any dictionary used as input must be static as well. ghstack-source-id: 248b4232ec03c7d3d21e1a45b59a112bfc2edeee Pull Request resolved: #40885
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.
Stack from ghstack:
TracingState::setValueassociates a concrete IValue in the tracedprogram with a
Value*symbolic. Previously, the logic for howGenericDicts worked was special cased to only work for very simple cases
and silently eat other cases.
This PR generalizes the logic to reflect the same behavior as using
dictionaries on input: whenever we encounter a dictionary in the system,
we completely "burn in" all the keys into the graph, and then
recursively call
setValueon the associated value.This has the effect of requiring that any dictionary structure you are
creating in a traced program be of fixed structure, similar to how any
dictionary used as input must be static as well.
Differential Revision: D22342490