Attempt at progressively feeding the Session to bypass type checking in #69
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
…rom a tensor This is some prep work for string output types and tensor types that vary across the model outputs. For now, the supported types are just the basic numeric types. Since strings have to be copied out of a tensor, it only makes sense to have `String` be an output type, not `str`, hence the new type so that we can have more input types supported than output types.
Outputs aren't all the same type for a single model, so this allows extracting different types per tensor.
This approach allocates owned Strings for each element, which works, but stresses the allocator, and incurs unnecessary copying. Part of the complication stems from the limitation that in Rust, a field can't be a reference to another field in the same struct. This means that having a Vec<u8> of copied data, referred to by a Vec<&str>, which is then referred to by an ArrayView, requires a sequence of 3 structs to express. Building a Vec<String> gets rid of the references, but also loses the efficiency of 1 allocation with strs pointing into it.
Codecov Report
@@ Coverage Diff @@
## master #69 +/- ##
===========================================
+ Coverage 14.58% 44.66% +30.08%
===========================================
Files 18 19 +1
Lines 960 1097 +137
===========================================
+ Hits 140 490 +350
+ Misses 820 607 -213
Continue to review full report at Codecov.
|
|
Sorry for the delay; I've been busy with other projects but I'll be able to take a look soon (famous last words...). |
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Outdated suggestions cannot be applied.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.
@marshallpierce I would love some feedback about this approach for differently typed tensors for inputs.
Just as a reminder,the inputs of the graph I have contain differently shaped and differently typed arrays
(input_ids, i64, [x, y]) , (past_key_values, f32, (x, y, z, m))
The session starts keeping the pointers, which means I have to leak the pointer from the
feedfunction, and drop them after running the session.The good thing with this approach is that
runcan keep backward compatibility and still enable more advanced usages (and no super complex types and extractions.