[None][feat] cache transceiver test in Perf sanity#16674

Open

chuangz0 wants to merge 19 commits into

NVIDIA:mainfrom

chuangz0:perf_sanity_cache_transceiver_test

Commits on Jul 24, 2026

[None][test] Add cache transceiver precheck driver for disagg perf sanity
chuangz0
committed
[None][infra] Gate disagg perf sanity stages on a cache transceiver precheck
chuangz0
committed
[None][infra] Print the test-results artifact URL when a Slurm stage fails
chuangz0
committed
[None][test] Drop pickle from the precheck ZMQ control channel
chuangz0
committed
[None][test] Support KVCacheManagerV2 in the cache transceiver precheck
chuangz0
committed
[None][test] Guard the precheck rendezvous against stale files from reused work dirs
chuangz0
committed
[None][test] Keep ctx precheck instances alive until the gen finishes all peers
chuangz0
committed
[None][test] Add stall-debugging logs to the precheck driver
chuangz0
committed
[None][test] Budget NIXL agent wire-up in the precheck's first rep
chuangz0
committed
[None][test] Simplify the precheck: one owner per policy, leaner hot path
chuangz0
committed
[None][test] Readability pass on the precheck driver
chuangz0
committed
[None][test] Stop capturing NIXL_* env in precheck verdicts
chuangz0
committed
[None][test] Run the network check on unmodelable models via a simplified pool; add Python-transceiver bandwidth
chuangz0
committed
[None][test] Rename the precheck's 'chunk' concept to 'wave'
chuangz0
committed
[None][test] Fail-fast the cache-transceiver precheck and fix review findings
chuangz0
committed
[None][test] Fix pre-commit lint findings across the precheck code
chuangz0
committed
[None][infra] Print the exact test-results artifact URL on Slurm failure
chuangz0
committed
[None][fix] Address review: MTP depth fallback, abort-flag cleanup and collective read
chuangz0
committed
[None][test] Add unit, contract and single-GPU e2e coverage for the cache-transceiver precheck driver
chuangz0
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[None][feat] cache transceiver test in Perf sanity#16674

[None][feat] cache transceiver test in Perf sanity#16674
chuangz0 wants to merge 19 commits into
NVIDIA:mainfrom
chuangz0:perf_sanity_cache_transceiver_test

Commits on Jul 24, 2026

[None][test] Add cache transceiver precheck driver for disagg perf sanity

[None][infra] Gate disagg perf sanity stages on a cache transceiver precheck

[None][infra] Print the test-results artifact URL when a Slurm stage fails

[None][test] Drop pickle from the precheck ZMQ control channel

[None][test] Support KVCacheManagerV2 in the cache transceiver precheck

[None][test] Guard the precheck rendezvous against stale files from reused work dirs

[None][test] Keep ctx precheck instances alive until the gen finishes all peers

[None][test] Add stall-debugging logs to the precheck driver

[None][test] Budget NIXL agent wire-up in the precheck's first rep

[None][test] Simplify the precheck: one owner per policy, leaner hot path

[None][test] Readability pass on the precheck driver

[None][test] Stop capturing NIXL_* env in precheck verdicts

[None][test] Run the network check on unmodelable models via a simplified pool; add Python-transceiver bandwidth

[None][test] Rename the precheck's 'chunk' concept to 'wave'

[None][test] Fail-fast the cache-transceiver precheck and fix review findings

[None][test] Fix pre-commit lint findings across the precheck code

[None][infra] Print the exact test-results artifact URL on Slurm failure

[None][fix] Address review: MTP depth fallback, abort-flag cleanup and collective read

[None][test] Add unit, contract and single-GPU e2e coverage for the cache-transceiver precheck driver