docs(recipe): add DeepSeek-V4-Flash Async CAM recipe

[ShwStone]

Purpose

Add a user-facing DeepSeek-V4-Flash W8A8 deployment recipe for the
CAMAsyncAFDConnector performance work tracked by #227.

The recipe describes a single-node, 16-NPU deployment:

• Attention: DP4 × TP2 on devices 0-7, with FlashComm1/SP enabled;
• FFN: DP8 × TP1 / EP8 on devices 8-15, with FlashComm1 disabled;
• AFD-managed two-stage token split;
• MBT 65536, shared compressor workspace, and a 4096 MB CAM HCCL buffer.

Issue

• Related issue: [DeepSeek V4][NPU] Support CAMAsyncAFDConnector #227

Scope

In scope

• Add a portable launch.sh <attention|ffn> entry point.
• Document runtime dependencies, topology, configuration, launch commands,
  limitations, and readiness checks.
• Record the corresponding issue [DeepSeek V4][NPU] Support CAMAsyncAFDConnector #227 performance summary.
• Link the recipe from the recipe index and CAM Async user guide.

Out of scope

• Runtime, connector, model, scheduler, or kernel changes.
• Multi-node orchestration or environment-specific job tooling.
• New NPU qualification runs.

Runtime dependency

The recorded performance cell used:

• vLLM 568afb3a1;
• vLLM-Ascend 80d8c194f;
• commit 2e01d4c5 from the still-open
  vLLM-Ascend PR #15452;
• follow-up slot-mapping fix e19e14da;
• CANN 9.0.1 and CAM 209.x operators.

PR #15452 provides the DeepSeek-V4 compressor-tail workspace reuse required
for the MBT 65536 configuration. The recipe explicitly documents that this
support is not yet available from the vLLM-Ascend base commit alone.

Implementation notes

The launcher:

• accepts attention or ffn as a positional role;
• requires only the model path, local communication IP, and network interface;
• contains no cluster-specific paths, IP addresses, or orchestration commands;
• configures FlashComm1 only for Attention;
• verifies the model configuration and CAM op-api library;
• handles CAM 209.x installations that expose only libcust_opapi.so;
• does not impose a startup delay or role ordering;
• does not enable vLLM native DBO.

Recorded performance context

For the issue #227 single-node DP4TP2/EP8 token-split matrix, all offered-rate
cells completed 512/512 requests without failure.

• 1.0x: 32,830 effective input tokens/s, TTFT p99 15.6s;
• 1.5x: 40,096 effective input tokens/s, TTFT p99 34.3s;
• versus the best baseline point, effective throughput improved by 4.8% at
  1.0x, 6.6% at 1.25x, and 8.6% at 1.5x;
• at 1.0x, TTFT p50 improved by 37.4%, TTFT p99 by 22.1%, and the 10-second
  TTFT SLO rate increased from 65.2% to 85.2%.

These are historical results for the recorded software and workload cell, not
a performance guarantee or a new qualification result for this PR.

Test result

• bash -n recipe/npu/CAMAsyncAFDConnector/deepseek_v4_flash/launch.sh — passed
• git diff --check — passed
• NPU execution — not run
• ShellCheck — not run because shellcheck is unavailable in the local environment

Docs impact

Updated:

• recipe/README.md
• docs/npu/CAM_ASYNC_CONNECTOR_USER_GUIDE.md

Added:

• recipe/npu/CAMAsyncAFDConnector/deepseek_v4_flash/README.md
• recipe/npu/CAMAsyncAFDConnector/deepseek_v4_flash/launch.sh

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.