[Engine] Allow manually aborting reload, fix unexpected deviceLostError #525
Conversation
This file contains hidden or 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
7 tasks
Member
|
Currently blocked by unexpected deviceLostError. Will fix that before merging this PR. |
Member
|
Depends on apache/tvm#17250 |
CharlieFRuan
changed the title
CharlieFRuan
added a commit
that referenced
this pull request
Aug 8, 2024
No breaking changes. The only diff is the following PR: - #525 - This PR updates the engine reload() and unload() methods to allow users to abort an uncompleted reload() by either: - call unload() any time before reload() completed - call reload() again before the previous reload() completed - Besides, it fixes the previous issue where `device lost error` is raised unexpectedly when user simply switches a model ### TVMjs - To support the above PR, TVMjs is updated and compiled at apache/tvm@1fcb620 - Difference: - Device error lost fix: apache/tvm#17250 - Add AbortSignal to fetching APIs: - apache/tvm#17208 - apache/tvm#17227 - apache/tvm#17233
jingyi-zhao-01
pushed a commit
to jingyi-zhao-01/web-llm
that referenced
this pull request
Dec 8, 2024
…or (mlc-ai#525) ### Manually aborting reload This PR updates the engine `reload()` and `unload()` methods to allow users to abort an uncompleted `reload()` by either: - call `unload()` any time before `reload()` completed - call `reload()` again before the previous `reload()` completed ### Note on unload() and unexpected device lost error Previously, we had an issue where a device lost error is reported when we simply switch a model intentionally (i.e. calling `reload()`). This is because `unload()` sets `deviceLostIsError` back to true immediately after calling `this.pipeline.dispose()`, which destroys the WebGPU device internally. However, WebGPU is asynchronous and may not finish after `dispose()` returns. This PR also fixes this issue by making `unload()` wait until the device is actually destroyed by introducing `LLMChatPipeline.sync()`. --------- Co-authored-by: Charlie Ruan <[email protected]>
jingyi-zhao-01
pushed a commit
to jingyi-zhao-01/web-llm
that referenced
this pull request
Dec 8, 2024
No breaking changes. The only diff is the following PR: - mlc-ai#525 - This PR updates the engine reload() and unload() methods to allow users to abort an uncompleted reload() by either: - call unload() any time before reload() completed - call reload() again before the previous reload() completed - Besides, it fixes the previous issue where `device lost error` is raised unexpectedly when user simply switches a model ### TVMjs - To support the above PR, TVMjs is updated and compiled at apache/tvm@1fcb620 - Difference: - Device error lost fix: apache/tvm#17250 - Add AbortSignal to fetching APIs: - apache/tvm#17208 - apache/tvm#17227 - apache/tvm#17233
atebites-hub
pushed a commit
to atebites-hub/web-llm
that referenced
this pull request
Oct 4, 2025
…or (mlc-ai#525) ### Manually aborting reload This PR updates the engine `reload()` and `unload()` methods to allow users to abort an uncompleted `reload()` by either: - call `unload()` any time before `reload()` completed - call `reload()` again before the previous `reload()` completed ### Note on unload() and unexpected device lost error Previously, we had an issue where a device lost error is reported when we simply switch a model intentionally (i.e. calling `reload()`). This is because `unload()` sets `deviceLostIsError` back to true immediately after calling `this.pipeline.dispose()`, which destroys the WebGPU device internally. However, WebGPU is asynchronous and may not finish after `dispose()` returns. This PR also fixes this issue by making `unload()` wait until the device is actually destroyed by introducing `LLMChatPipeline.sync()`. --------- Co-authored-by: Charlie Ruan <[email protected]>
atebites-hub
pushed a commit
to atebites-hub/web-llm
that referenced
this pull request
Oct 4, 2025
No breaking changes. The only diff is the following PR: - mlc-ai#525 - This PR updates the engine reload() and unload() methods to allow users to abort an uncompleted reload() by either: - call unload() any time before reload() completed - call reload() again before the previous reload() completed - Besides, it fixes the previous issue where `device lost error` is raised unexpectedly when user simply switches a model ### TVMjs - To support the above PR, TVMjs is updated and compiled at apache/tvm@1fcb620 - Difference: - Device error lost fix: apache/tvm#17250 - Add AbortSignal to fetching APIs: - apache/tvm#17208 - apache/tvm#17227 - apache/tvm#17233
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.
Manually aborting reload
This PR updates the engine
reload()andunload()methods to allow users to abort an uncompletedreload()by either:unload()any time beforereload()completedreload()again before the previousreload()completedExample added in
examples/abort-reload.Console output:
Related issues:
#484
#499
Note on unload() and unexpected device lost error
Previously, we had an issue where a device lost error is reported when we simply switch a model intentionally (i.e. calling
reload()). This is becauseunload()setsdeviceLostIsErrorback to true immediately after callingthis.pipeline.dispose(), which destroys the WebGPU device internally. However, WebGPU is asynchronous and may not finish afterdispose()returns. This PR also fixes this issue by makingunload()wait until the device is actually destroyed by introducingLLMChatPipeline.sync(). Otherwise, the old device's device-lost callback may still be triggered. See apache/tvm#17250