Add DCP (Decode Context Parallelism) support for the FlashInfer backend:
- allocator.py: DCP interleaved token allocation with rank-based filtering
- forward_batch_info.py: missing import for create_chunked_prefix_cache_kv_indices
- deepseek_v2.py: FORWARD_ABSORB_CORE_ATTENTION_BACKENDS, merge_state_v2 import
- scheduler.py: DCP rank in process title

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…allelism

Add DCP support for the FlashAttention (FA3) backend:
- FlashAttentionMetadata: dcp_page_table and dcp_cache_seqlens for local KV shard
- _init_dcp_decode_metadata: vectorized page table construction for DCP interleaving
- Forward decode: DCP AllGather Q, local page table substitution, LSE normalization
- CUDA graph: pre-allocated DCP buffers for graph capture/replay
- deepseek_v2.py: is_base_e LSE detection and ln(2) conversion for AG+RS

Add All-to-All (A2A) as an alternative to AG+RS for DCP communication:
- pynccl.py: all_to_all_single using ncclGroupStart/End for graph capture
- parallel_state.py: GroupCoordinator.all_to_all_single method
- dcp_a2a.py: Triton LSE-weighted combine kernel, fused output+LSE buffer
- server_args.py: --dcp-comm-backend {ag_rs, a2a} flag
- deepseek_v2.py: A2A dispatch in DCP reduce block
- flashattention_backend.py: A2A dispatch, per-bs CUDA graph buffer allocation
- Tests: test_dcp_a2a.py (Triton kernel correctness), test_dcp_accuracy_matrix.py

Replace the SGLANG_DCP environment variable with a proper --dcp-size
argument in ServerArgs for decode context parallelism configuration:
- server_args.py: add dcp_size field and --dcp-size argparse argument
- parallel_state.py: add set_dcp_size(), remove env var read from
  get_dcp_size_from_env(), pass dcp_size through initialize_model_parallel
- model_runner.py: pass server_args.dcp_size to initialize_model_parallel
- scheduler.py: use server_args.dcp_size instead of env var
- engine.py: use server_args.dcp_size instead of env var
- test_dcp_accuracy_matrix.py: use --dcp-size flag instead of env var

CPTritonContext was intended to cache compiled Triton kernels across
calls, but ctx was never persisted — a new instance was created on
every call, so the caching never took effect. Triton already caches
compiled kernels internally based on constexpr values, making this
wrapper redundant. Remove it and call the kernel directly.

…ersion hack

The AG+RS kernel (_correct_attn_cp_out_kernel) was hardcoded to base-2
(exp2/log2), requiring callers to manually convert base-e LSE from FA3
via division by ln(2). This was error-prone — flashattention_backend.py
was missing the conversion entirely for its non-MLA AG+RS path.

Add IS_BASE_E constexpr to the kernel so it handles both base-e (exp/log)
and base-2 (exp2/log2) natively, matching _dcp_lse_combine_kernel's
approach. Propagate is_lse_base_on_e through correct_attn_out and
cp_lse_ag_out_rs, and remove the lse / ln(2) hack from deepseek_v2.py.

Remove TODO comments from the original DCP commit (3c751c6) where the
described functionality has since been implemented:
- flashinfer_mla_backend.py: LSE return for decode (done), kv_indices
  update (filter_seq_indices implemented below the TODO)
- deepseek_v2.py: change local_heads logic (attn_mqa_for_dcp_decode
  created), all_gather q_pe (implemented below), return lse and correct
  attn_output (DCP reduce block implemented)
- model_runner.py: prepare for dcp (DCP prefix preparation implemented)
- allocator.py: triton kernel for filter_local_indices (only used in
  offline save/load paths, PyTorch boolean indexing is sufficient)

Register DCP tests in test/registered/ for CI auto-discovery:
- dcp/test_dcp_accuracy.py: E2E 8-GPU GSM8K accuracy (FlashInfer/FA3 x AG+RS/A2A)
- kernels/test_dcp_lse_combine.py: Triton LSE combine vs CPU reference
- kernels/test_dcp_interleaved.py: KV allocator interleaved storage
- kernels/test_dcp_fa3_standalone.py: FA3 MLA simulated DCP sharding
- unit/server_args/test_dcp_config.py: ServerArgs DCP validation
- unit/layers/test_dcp_cascade_guard.py: cascade attention guard logic
- unit/layers/test_dcp_need_lse.py: need_lse flag logic

Made-with: Cursor

- bench_dcp_serving.sh: automated benchmark across 6 configs
  (TP8 baseline, DCP8 AG+RS, DCP8 A2A) x (FlashInfer, FA3) with
  accuracy (GSM8K) + throughput (bench_serving) at cc1-512
- plot_dcp_bench.py: parse bench_serving outputs, generate comparison
  charts and markdown tables for throughput/TTFT/TPOT/ITL metrics

Made-with: Cursor

The DCP cherry-pick accidentally duplicated all mixin methods inside
deepseek_v2.py (3455 vs 2264 lines). Reset deepseek_v2.py to main and
move DCP-specific code to the proper mixin files:

- forward_mla.py: Q AllGather (decode), KV prefix gather (extend),
  LSE-weighted combine (AG+RS and A2A) in forward_absorb_prepare/core
- forward_mha.py: DCP KV index filtering in _get_mla_kv_buffer, DCP
  prefix KV AllGather in forward_normal_prepare and _chunked_prefix_attn_mha
- deepseek_v2.py: only DCP imports, attn_mqa_for_dcp_decode init,
  and _all_gather_dcp_kv_cache method
- utils.py: add_forward_absorb_core_attention_backend helper

Made-with: Cursor

Allocate A2A send/recv buffers inside SymmetricMemoryContext so they
are graph-capturable with PyNccl. Benchmark shows 3.0-3.6x speedup
for A2A collectives with symmetric memory vs torch eager.

Remove --disable-cuda-graph from A2A test configs since A2A now
works with CUDA graph capture.

Made-with: Cursor

Use explicit `shift 1` to make intent clear -- shift only the first
arg (NGPU), leaving remaining args as OPS. The `2>/dev/null` is stderr
redirection, not a shift count. Address review feedback from PR sgl-project#21637.

Made-with: Cursor

- Fix CPU-device torch.arange in _all_gather_dcp_kv_cache causing
  CPU-GPU sync on every call (add device=kv_a.device)
- Add assertion H % N == 0 in dcp_a2a_lse_reduce to guard against
  silent wrong results with indivisible head counts
- Clarify return statement in dcp_lse_combine_triton with explicit
  parentheses for operator precedence
- Remove wasteful tl.load in Triton kernel store -- use out_ptr.dtype
  instead of loading from recv_output_ptr just to get the dtype

Made-with: Cursor