[sglang-miles] Sync #40648: drop the DP-attention NCCL graph registration trigger

[guapisolo]

Summary

Stop disabling NCCL graph registration for DP-attention engines whose graph pool is not pausable.

Symptom & Reproduction

• Symptom: every sglang-miles engine with --enable-dp-attention, --dp-size > 1 and decode CUDA graphs defaults to NCCL_GRAPH_REGISTER=0 even without --enable-memory-saver, so its graph-captured collectives lose NCCL zero-copy registration.
• Reproduction: on sglang-miles 6e25d1bd57, test_without_all_to_all_env_is_untouched resolves a DP-attention config (tp_size=2, dp_size=2, default decode graphs) and expects NCCL_GRAPH_REGISTER to stay unset; it only holds because the test stubs _dp_attention_replays_decode_graphs to False. Without the stub, handle_data_parallelism sets it to 0.

Root Cause

1. #40977 picked #40648 at 0854ad8c52, before review removed its DP-attention trigger.
2. _dp_attention_replays_decode_graphs fires for every DP-attention engine that replays decode graphs.
3. #40648 review found no hang from registration on a non-pausable pool, so upstream dropped it.

Fix

Two commits, so the swap is reviewable:

• 616a8d299b reverts only the fix(nccl): disable graph buffer registration for pausable graph pools #40648 hunks of [sglang-miles] Cherry-pick #40448, #40648, #40978, #40979, #40980 for MiMo-V2.6-Flash colocated RL #40977, restoring parallel_hook.py and test_server_args.py to their pre-[sglang-miles] Cherry-pick #40448, #40648, #40978, #40979, #40980 for MiMo-V2.6-Flash colocated RL #40977 bytes; the other four picks are untouched.
• b35cf3e72c applies upstream fix(nccl): disable graph buffer registration for pausable graph pools #40648 at 8739ae8063: only the pausable graph pool trigger (--enable-memory-saver with SGLANG_MEMORY_SAVER_CUDA_GRAPH) remains, and the docs describe the hazard as stale capture-time registrations after a torch_memory_saver resume (NCCL_GRAPH_REGISTER=1 can hang after a same-VA TMS resume fzyzcjy/torch_memory_saver#88).

Colocated miles engines (offload rollout plus miles' default SGLANG_MEMORY_SAVER_CUDA_GRAPH=true) still get NCCL_GRAPH_REGISTER=0 with no miles change; DP-attention engines without a pausable pool return to NCCL's default of 1. An explicit NCCL_GRAPH_REGISTER still wins, with a warning.

Verification

• git diff: after 616a8d299b both files equal sglang-miles before [sglang-miles] Cherry-pick #40448, #40648, #40978, #40979, #40980 for MiMo-V2.6-Flash colocated RL #40977; at this head the guard region of parallel_hook.py and TestTpLmHeadAllToAllNcclGraphRegister are byte-identical to upstream fix(nccl): disable graph buffer registration for pausable graph pools #40648 at 8739ae8063.
• ruff format --check, ruff check --select F, and py_compile on the changed files: pass.
• Not run on this head: test_server_args.py -k NcclGraphRegister and GPU runs; the authoring host has no torch.

Review Focus

• _graph_pool_is_pausable / parallel_hook.py: the remaining trigger must match the graph backends' enable_memory_saver and SGLANG_MEMORY_SAVER_CUDA_GRAPH gate.
• Behavior change on sglang-miles: DP-attention engines with a non-pausable pool regain NCCL graph registration.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #36804707958
Latest PR Test (Extra): ❌ Run #36804707751
Latest PR Test (AMD ROCm 10): ❌ Run #36804707979