Skip to content

[pull] master from tensorflow:master#536

Merged
pull[bot] merged 24 commits intonoaai:masterfrom
tensorflow:master
May 23, 2025
Merged

[pull] master from tensorflow:master#536
pull[bot] merged 24 commits intonoaai:masterfrom
tensorflow:master

Conversation

@pull
Copy link
Copy Markdown

@pull pull bot commented May 23, 2025

See Commits and Changes for more details.


Created by pull[bot] (v2.0.0-alpha.1)

Can you help keep this open source service alive? 💖 Please sponsor : )

vsytch and others added 24 commits May 23, 2025 10:08
It is beneficial to create artifical scheduling boundaries via called computations and then inline afterwards, as schedulers only act on a single computation at a time.

PiperOrigin-RevId: 762467346
…es the crash in jax B200 test, that was caused by recent triton integration.

PiperOrigin-RevId: 762472690
This is a cleanup. In the following changes, I want to add move dtype+reduction_kind pairs to the kernel and the amount of templating that needs to be done in this test to support different combination of reductions is getting out of hand.

PiperOrigin-RevId: 762477599
create a fixture that gets the registered plugins, asserts there's only one, and then initializes it as part of the fixture.

PiperOrigin-RevId: 762488288
Keeping each loop iteration as a callable simplifies scheduling drastically.

PiperOrigin-RevId: 762493037
… move it to being updated only when the graph is initialized/changed and fetch from storage in the hot path.

Tidy up 'Does(Occupy/Free)Resources' by computing during initialization, getting rid of std::optional / lazy init.

PiperOrigin-RevId: 762495390
Skip expanding already ignored nodes when adding nodes to the ignored set.
Skip expanding nodes mapped by the same matcher when mapping the subgraph.

PiperOrigin-RevId: 762496848
… repacked allocations import logic to build the buffer intervals map in a `forward` manner.

Previously, we used to loop over the imported allocations, and check if a collocation of it has already been added as primary (need_allocation=true) to the full_buffer_interval_map_. Instead, we now jointly import the primary and all its colocations to full_buffer_interval_map_.

PiperOrigin-RevId: 762521615
PiperOrigin-RevId: 762534581
The issue here is that the tests generate random input sizes, and then compute the output size. The problem is that if the stride of the kernel is > 1, then some input sizes not aligned to the kernel stride will result in area in the input tensor that is not actually used by the convolution result, but *is* used by the dynamic quantization range computation.

PiperOrigin-RevId: 762536789
…FromProto(ShapeProto)`

The old constructor is deprecated.

PiperOrigin-RevId: 762537880
PiperOrigin-RevId: 762548187
Specifically, VLOG(1) --> VLOG(3) and VLOG(2) -> VLOG(4).

PiperOrigin-RevId: 762568435
…tation and workflow files for consistency

PiperOrigin-RevId: 762570935
PiperOrigin-RevId: 762572277
…ue to allow insecure credentials.

PiperOrigin-RevId: 762577138
… too big to fit uint16_t type of op_params' version

PiperOrigin-RevId: 762584396
PiperOrigin-RevId: 762588691
@pull pull bot added the ⤵️ pull label May 23, 2025
@pull pull bot merged commit bc1fbcf into noaai:master May 23, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.