Skip to content

Tags: google/heir

Tags

nightly-2026.08.16

Toggle nightly-2026.08.16's commit message
Integrate LLVM at llvm/llvm-project@75a45c373407

Updates LLVM usage to match
[75a45c373407](llvm/llvm-project@75a45c373407)

PiperOrigin-RevId: 964904972

nightly-2026.08.15

Toggle nightly-2026.08.15's commit message
Integrate LLVM at llvm/llvm-project@75a45c373407

Updates LLVM usage to match
[75a45c373407](llvm/llvm-project@75a45c373407)

PiperOrigin-RevId: 964904972

nightly

Toggle nightly's commit message
Integrate LLVM at llvm/llvm-project@75a45c373407

Updates LLVM usage to match
[75a45c373407](llvm/llvm-project@75a45c373407)

PiperOrigin-RevId: 964904972

nightly-2026.08.14

Toggle nightly-2026.08.14's commit message
Set grpc_keepalive_time to help prevent BuildBuddy disconnects

I still see timeouts on PRs and then the GHA disconnects from the buildbuddy
server. E.g.,

  https://heir.buildbuddy.io/invocation/ebd5b0de-618e-4419-b4ef-94b4b573b0d4

for this job:

  https://github.com/google/heir/actions/runs/31540176730/job/93940562317

After consulting with BuildBuddy support, they know that networking issues on
GitHub or the AWS/Azure side (whatever cloud is hosting their runners) can
cause this, and recommended we set grpc_keepalive_time to hedge against that.

They also mentioned upgrading to bazel 9 would help, though that is currently
blocked by #3113 (comment)
for at least one more month.

PiperOrigin-RevId: 964161074

nightly-2026.08.13

Toggle nightly-2026.08.13's commit message
Merge pull request #3323 from google:mdgrs/layoutPropConv2dFixes

PiperOrigin-RevId: 963613808

v2026.08.11.dev0

Toggle v2026.08.11.dev0's commit message
Implement early bootstrap placement in HEIR

This adds a new pass `EarlyBootstrapPlacement` pass, specific to the torch pipeline, that does early greedy bootstrap insertion while linalg and chebyshev ops are still a single operation.

This prevents `SecretInsertMgmt` from inserting bootstraps inside loops or in bulk on loop outputs, resolving some performance issues we saw with the hotword model.

It's also a bit of a cop out for the ineptitude of the greedy bootstrap pass to handle loops, which I feel is later going to be resolved by switching to Edward's ILP-based mgmt solver.

This change also includes some of the hoisting logic from my attempts to get the `SecretInsertMgmt` to move unnecessary bootstraps out of loops, and I feel that for the timing with the upcoming blog post, I'd like to just merge this as techdebt and worry about cleaning it up later (or not, if we switch to the ILP solver).

PiperOrigin-RevId: 962470276

nightly-2026.08.12

Toggle nightly-2026.08.12's commit message
fix: implement physical slot replication for bicyclic and tricyclic l…

…ayout after

matrix multiplication.

In the bicyclic layout, we require for each 0 <= slot < numSlot, (ct, slot) is
mapped to (slot % n, slot % m). However, after computing through BSGS we can only
guarantee for all 0 <= slot < n*m this property preserves (indeed we can guarantee
more, but for the tail part it is not correct if n*m does not divide numSlot).

In this commit, we add a new relation called `periodic replication relation`
that replicate the first copy to all slots periodically. After each matrix
multiplication with CRT layouts (bicyclic, tricyclic), we compose this relation
to derive the result.

The cost of the layout switching will be up to logN rotations.

PiperOrigin-RevId: 962982741

nightly-2026.08.11

Toggle nightly-2026.08.11's commit message
Implement early bootstrap placement in HEIR

This adds a new pass `EarlyBootstrapPlacement` pass, specific to the torch pipeline, that does early greedy bootstrap insertion while linalg and chebyshev ops are still a single operation.

This prevents `SecretInsertMgmt` from inserting bootstraps inside loops or in bulk on loop outputs, resolving some performance issues we saw with the hotword model.

It's also a bit of a cop out for the ineptitude of the greedy bootstrap pass to handle loops, which I feel is later going to be resolved by switching to Edward's ILP-based mgmt solver.

This change also includes some of the hoisting logic from my attempts to get the `SecretInsertMgmt` to move unnecessary bootstraps out of loops, and I feel that for the timing with the upcoming blog post, I'd like to just merge this as techdebt and worry about cleaning it up later (or not, if we switch to the ILP solver).

PiperOrigin-RevId: 962470276

nightly-2026.08.10

Toggle nightly-2026.08.10's commit message
Integrate LLVM at llvm/llvm-project@a6b0af7536ef

Updates LLVM usage to match
[a6b0af7536ef](llvm/llvm-project@a6b0af7536ef)

PiperOrigin-RevId: 961806049

nightly-2026.08.09

Toggle nightly-2026.08.09's commit message
polynomial approx: implement square and multiply for math.fpowi

PiperOrigin-RevId: 961260919