feat: support decode context parallelism for PyTorch MLA

[CUHKSZzxy]

Motivation

Reduce replicated MLA KV-cache storage within attention tensor parallelism by adding opt-in decode context parallelism (DCP) to the PyTorch backend. DCP reuses existing TP ranks instead of increasing the model world size.

Modification

• Add --dcp / PytorchEngineConfig(dcp=...), DCP process groups, interleaved cache ownership, and scheduler accounting for logical blocks spanning multiple ranks.
• Support dense FlashMLA and sparse DSA attention with rank-local KV reads, query-head gathering, and LSE-based output merging. Sparse DSA exchanges and merges local top-k candidates into global selections.
• Support cached-prefix prefill through bounded MLA context gathering and FP32 accumulation across chunks. Reserve gather/merge and top-k workspace, including MTP-expanded verification rows.
• Support deepseek_mtp for dense and sparse MLA with per-query causal lengths and CUDA-graph-compatible metadata.
• Document the supported configurations and extend numerical/cache-layout tests, including BF16/FP8 cached prefill with DCP2/DCP4.

dcp=1 remains the default. TurboMind's existing --cp option is unchanged.

Supported scope and remaining work

• CUDA Hopper/SM90, FlashMLA-backed MLA, BF16 activations, DP1/EP1. FP8 model weights are compatible with BF16 activations; sparse MLA also supports the blocked-FP8 MLA KV layout.
• Sparse DSA top-k is limited to 512/2048 and requires compatible DeepGEMM, TileLang, and FlashMLA kernels.
• No sliding-window attention, MemDecode, prefill/decode disaggregation, or external KV-cache connectors in this initial scope.
• Flattened indexer KV relies on remaining memory headroom; the reservation is not a hard cap on all indexer temporary allocations. Decode-score chunking is not implemented.
• Keep this PR in draft while reviewing prefix-cache behavior and performance. Full-model cold/warm exact-output differences also reproduced with DCP disabled and with BF16 KV; no DCP-specific cache corruption was identified, but the source remains unisolated. This is not an unconditional prefix-cache correctness sign-off or a short-context speedup claim.

Validation

On the published head:

• pytest tests/pytorch/engine/test_executor_base.py tests/pytorch/test_dcp.py -q: 53 passed on CPU.
• pytest tests/pytorch/kernel/test_mla_attention.py -k dcp -q: 14 passed on one H200 GPU. Cached-prefill tests execute real cache-writing and attention kernels against an independent reference; DCP collectives are simulated in these single-GPU tests.
• All applicable pre-commit hooks passed.

Earlier full-model validation during branch development used GLM-5.2-FP8, TP8/DCP4, MTP5, and CUDA graphs. GSM8K scored 1268/1319 (96.1334%) with prefix caching off and 1266/1319 (95.9818%) with it enabled. The prefix-enabled dataset run had zero cache hits, so these scores do not validate reused-prefix correctness. Separate full-model replay controls exercised actual cache hits and exposed the parity caveat above. These were single runs, not a statistically established accuracy or performance comparison, and were not rerun for the final accounting-only change.

The full downstream dependency-version matrix has not been validated.

Assistance

Assisted with Codex + GPT-6-Astra High, reviewed manually