CollectiveX: move to NCCL EP v0.2 and extend platform coverage

[artpol84]

Summary

NCCL EP v0.2 fixes several critical bugs that were limiting the set of supported platforms for LL mode:

• Support for multiple handles per group was fixed (NVIDIA/nccl-extensions@825d722)
• The combine pipeline was properly fenced (NVIDIA/nccl-extensions@27e9533)

The release also resolves the NCCL dependency issue
that was causing HT IMA.

This commit retargets the nccl-ep backend to NCCL EP v0.2 and lifts
the previous platform and configuration limitations accordingly.

Thank you for considering our contribution,
NCCL EP team

Testing

GB200/GB300 platforms

Production sweep (nccl-ep, EP8 + EP16, gb200 + gb300 — 4 jobs, 22 cases total)
Full configs/sweep.json production profile (256 trials × 8 iters/trial, 32 warmup/trial, 128×4 chain sampling), generated via sweep_matrix.py/runtime/config.py (not hand-built argv). Covers normal mode (decode ladder 1–512 + prefill ladder 1024–8192) and low-latency mode (decode, _LL_LADDER_CAP-bounded) on both SKUs at both EP sizes.
Result: 22/22 cases status: success, cross_rank_consistent: true, zero correctness failures.

LL-fix extended verification (low-latency mode only, EP8 + EP16, gb200 + gb300 — 4 jobs, 4 cases)
Targeted re-test of the _LL_LADDER_CAP 128→256 fix at 10x the standard trial count (2560 trials/point vs. 256) across the full T=1–256 decode ladder, including T=256 — the exact point the pre-fix race was previously observed on (gb300 EP8, "1 failure in 5 executions") — and gb300 EP16, a case that was unmeasurable at all under the old T≤128 clamp.
Result: 4/4 cases status: success, cross_rank_consistent: true, zero failed token points across ~10k total trial-points measured.

Both rounds ran on the same container/backend-prep path (runtime/prepare_backend.sh unmodified; pinned nccl4py==0.3.1.post1/nccl-extensions==0.1.0).

B200

Production sweep (nccl-ep, EP8 + EP16, B200 — 2 jobs, 6 cases total)

Full configs/sweep.json production profile (256 trials × 8 iters/trial, 32 warmup/trial, 128×4 chain sampling), generated via sweep_matrix.py/runtime/config.py (not hand-built argv). Covers normal mode (decode ladder 1–512 + prefill ladder 1024–8192) and low-latency mode (decode, _LL_LADDER_CAP-bounded at the pre-change value of 128) at both EP sizes.

Result: 6/6 cases status: success, cross_rank_consistent: true, zero correctness failures (44/44 individual ladder points passed chain_last_output_passed + post_chain_state_passed, max relative error 0.0).

---

Note

Medium Risk
Changes benchmark dependency pins, multi-node NCCL EP coverage, and low-latency ladder sizing on real GPU fleets; correctness risk is mitigated by v0.2 fixes but sweep matrices and comparability across historical runs shift.

Overview
CollectiveX’s nccl-ep backend now installs nccl-extensions[cu13]==0.1.0 (NCCL EP v0.2) instead of the older nccl4py / contrib/nccl_ep path, with matching updates in runtime/common.sh and prepare_backend.sh (cache key, pip install, docs).

Coverage and measurement behavior change with that pin: platform_config.json enables nccl-ep EP16 in normal mode on H100/H200/B300 and low-latency EP16 on B200 (nscale), GB200, and GB300 where it was previously EP8-only or blocked. In ep_nccl.py, low-latency decode _LL_LADDER_CAP goes 128 → 256, dropping the work-around for the unfixed combine fence; docs/README/methodology now describe x86 EP16 GIN scale-out as supported for HT and BF16-only sweeps (v0.2 FP8 dispatch left as future work).

^{Reviewed by Cursor Bugbot for commit eca48f7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit f8fefc0. Configure here.}

[Oseltamivir]

Thanks for the work, some of our clusters are down are we're talking to Nvidia SRE regarding getting them back. Will work on this over the weekend.

[artpol84]

@Oseltamivir I have just pushed few more changes.
Most of them are aiming to actualize the documentation and I will appreciate your help to make it right.
There is also one code change (COLLX_NCCL4PY_SPEC → COLLX_NCCL_EP_SPEC rename) that I haven't yet tested also due to resource maintenance. I will verify it and update here as soon as I can.