-
Notifications
You must be signed in to change notification settings - Fork 75
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[DO NOT MERGE] Add Database.dump_pruned #204
Closed
Closed
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
MLC local ci setup.
This PR adds CI for Windows and macOS building, which may take 90-100 mins. Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn>
…ai#139) This PR implements basic DTensor structure, and its printer/parser support
mlc-ai#172) … reduction This PR modifies the behavior of tir.Schedule.blockize, so that when the outer block after blockization does no effective reduction, the T.init part will be kept in the inner block. NOTE: unit tests regarding blockize may fail due to structural inequality. Co-authored-by: Bohan Hou <spectrometerh@gmail.com> Co-authored-by: spectrometerHBH <spectrometerh@gmail.com>
Problem after S3's fix: - cannot get producer outside current scope - cannot handle dynamic shape in ReverseComputeInliner Fix: - try to get leaf producer block outside current scope - ignore non-index var when substituting in ReverseComputeInliner's substituter
Both PadEinsum and cross-thread reduction lowering is using BufferReplacer as a symbol, in which case the linker links in a wrong way and causes very severe and hard-to-find error. This PR renames the buffer replacer on both sides.
This PR fixes a bug of recent enhancement in mlc-ai#176. Credit goes to Bohan. Co-authored-by: Bohan Hou <spectrometerh@gmail.com>
This PR is a reasonable but hot fix. Will send it to unity today with regression.
This PR aims to fix a relatively rare flaky race issuse when we are attempting to purge memory and the buffer is still in use in command buffer
…#194) This PR fixes an issue in MetaScheduleApplyDatabase, which uses the IRModule being queried as the base module of TIR Schedule. This will cause the inconsistency between the IRModule being scheduled and the trace of the record. For example, the trace of the record may have `get-block` instructions which tries to get block from a given name, which exists in the IRModule of the record while does not exist in the IRModule being queried (which is right the IRModule being scheduled). Therefore, this PR adds a case discussion. When anchor-op equality is not applied, we create the TIR schedule from the IRModule of the record, so that we can prevent the issue above from happening. Will send this PR to unity soon, together with unit tests.
This PR updates the search patter of config.cmake to CMAKE_BINARY_DIR. This behavior remains unchanged for TVM build. The difference only happens when tvm was included as a subdirectory by another project via add_subdirectory. In such cases, we will search the config.cmake on the root of the parent project rather than the tvm folder.
This PR add a smart mode of param fetching that can pass in -1 and fetches all parameters with given prefix.
This PR updates the metal runtime storage format to directly store the kernel map. This will enable more robust support to leverage metallib binary format which may not be compatible with previous string split. It changes the binary format of the metal module. We also added a version to enable easier future update.
This PR fixes the forece narrow to i32 pass so it ignores integer with smaller bitwidth
This PR add system lib build option to relax also adds an util that creates an archived static library in compilation
This PR bugfixes ring buffer ReadWithCallback when the callback send function read part of the data. Also backported one robustness fix from unity. Testcases are added
This PR makes CompositeFunctionAnnotator process all Relax functions rather than just the 'main' function. This enables running cutlass partitioning on LLaMA from web-llm.
junrushao
force-pushed
the
feature/2023-05-03/prune-db
branch
from
May 5, 2023 21:14
796b9c2
to
cf40773
Compare
junrushao
changed the title
[WIP] Add Database.dump_pruned
[DO NOT MERGE] Add Database.dump_pruned
May 6, 2023
Closing as all the changes have been committed to mainline |
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 PR has been decomposed into 5 and sent separately to mainline:
tune_tir
to tune IRModule of TIR Collections apache/tvm#14784