Derive CAM HCCL buffer sizes per role

[ShwStone]

Summary

• derive independent Attention and FFN CAM HCCL buffer sizes from the model, scheduler, topology, and quantization configuration
• apply 1.1x headroom through connector-scoped HCCL process-group options for both CAM Async and CAMP2P
• log each role's selected buffer at startup
• warn when the memory outside gpu_memory_utilization is less than 2.5x the role-local CAM buffer, without changing the configured utilization
• preserve the process-wide HCCL_BUFFSIZE setting for unrelated HCCL groups
• update the CAM Async and CAMP2P guides, launch recipes, and CPU-safe coverage

Motivation

HCCL_BUFFSIZE is process-wide, so increasing it for CAM also enlarges unrelated TP, DP, PCP, or EP communication buffers. The derived size is now scoped to connector-owned HCCL process groups. HCCL_BUFFSIZE is neither overwritten nor unset, so users can still tune other HCCL groups intentionally.

The derived requirements are:

tokens_per_npu = ceil(max_num_batched_tokens / num_npus_per_dp_group)
attention_bytes = 2 * hidden_size * tokens_per_npu * (topk + 1)
ffn_bytes = num_attention_ranks * (6176 if dynamic quantization else 12288) * tokens_per_npu
role_buffer_mb = ceil(1.1 * role_bytes / 1_MiB)

The FFN formula intentionally has no topk + 1 multiplier. num_npus_per_dp_group describes the total NPU count within one Attention DP group; the currently supported topology realizes that dimension with TP.

Validation

• uv run --locked ruff check .
• uv run --locked ruff format --check .
• uv run --locked pytest -q tests/unit -m 'not gpu and not vllm_runtime'
• python3 -m compileall -q afd_plugin tests
• bash -n recipe/npu/CAMAsyncAFDConnector/deepseek_v3_2/v0_26_accuracy/attention_dp2tp8.sh recipe/npu/CAMAsyncAFDConnector/deepseek_v3_2/v0_26_accuracy/ffn_ep16.sh recipe/npu/CAMP2pAFDConnector/deepseek_v3_2/afd_attention.sh recipe/npu/CAMP2pAFDConnector/deepseek_v3_2/afd_ffn.sh
• git diff --check

NPU end-to-end validation is still pending.

[ShwStone]

BYTES_PER_MB is kind of stupid

[ShwStone]

this hccl-buffer auto planning should not just use for Async. pls impl for both async & p2p