spectrometerHBH opened a new pull request, #20180:
URL: https://github.com/apache/tvm/pull/20180
## Motivation
Calling `K.cta_id_in_cluster(...)` is an explicit request for CUDA
cluster-launch semantics, including when every cluster extent is one. Lowering
previously removed extent-one `clusterCtaIdx.*` bindings, and the CUDA runtime
inferred cluster presence only from dimensions greater than one. As a result,
an explicit `(1, 1, 1)` cluster launch became indistinguishable from an
ordinary CTA launch.
## Changes
- Preserve `clusterCtaIdx.*` launch bindings whenever cluster CTA scope is
present, including extent-one bindings.
- Derive a `use_cluster_launch` flag from regular cluster launch-tag
presence in `LaunchParamConfig`.
- Emit `CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION` whenever that flag is set,
including `(1, 1, 1)`.
- Keep the missing-`clusterCtaIdx` fallback to constant zero for
compatibility with pre-existing IR.
- Add lowering, `SplitHostDevice`, and C++ launch-parameter coverage for
ordinary, singleton, multi-CTA, and preferred-cluster cases.
This does not change serialized module metadata. Preferred-cluster tags
alone do not request a regular cluster launch.
## Launch behavior
NCU 2026.1 measurements on NVIDIA B200:
| Probe | Cluster dimensions | Cluster Size | Uses Blocks as Clusters |
|---|---:|---:|---:|
| Ordinary CTA | `(0, 0, 0)` | 0 | 0 |
| Explicit singleton | `(1, 1, 1)` | 1 | 0 |
| Explicit multi-CTA | `(2, 1, 1)` | 2 | 0 |
## Validation
- TVM build completed successfully.
- LowerTIRx cluster/preferred-cluster tests: 6 passed.
- SplitHostDevice launch tests: 6 passed.
- `LaunchParamConfig` C++ test: 1 passed.
- Relevant CUDA cluster codegen tests: 2 passed.
- Downstream affected correctness matrix: 278/278 passed without skips:
- amax: 59
- fast top-k: 34
- dense SwiGLU: 75
- interleaved quant: 58
- grouped dGLU: 52
- Staged-file pre-commit checks passed.
Downstream B200 benchmarks used Proton timing, five independent samples per
implementation, one-second cooldowns, reference implementations enabled,
interference monitoring, and no sample splicing:
| Matrix | Passing rows | Minimum `mean(reference) / mean(tirx)` |
|---|---:|---:|
| Targeted affected-family gate | 15/15 | 0.996504 |
| Full amax matrix | 88/88 | 0.997440 |
| Full fast top-k matrix | 34/34 | 1.006541 |
The maximum patched-versus-main TIRx slowdown in the targeted comparison was
0.911%, below the 1% investigation threshold. Six externally interfered amax
attempts were recorded and replaced only by complete five-round row reruns.
A full TIRx suite run was also performed. The strict kernel registry passed.
The broad parallel run encountered unrelated environment/reference failures
from cross-device DLPack state, aggregate GPU memory pressure, a missing
optional `cutlass.experimental` package, and reference PTX incompatibility. An
isolated core rerun produced 2631 passed, 73 skipped, 3 xpassed, and one
pre-existing stale packed-dtype error-message assertion; no affected cluster
correctness configuration failed.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
