SIGN IN SIGN UP

[Doc][Feature] Add Atlas 800 A2 deployment for Qwen3.8-2.4T-A95B (#14896)

### What this PR does / why we need it?

Adds the validated eight-node Atlas 800 A2 (64GB × 8) mixed
Prefill/Decode
deployment (DP8/TP8/EP64, W4A8 checkpoint) to the Qwen3.8-2.4T-A95B
tutorial:

  - Docker installation and multi-node deployment tabs for the A2 series
(Node 0 / Nodes 1-7 startup commands, master-worker difference table,
    key deployment parameters table)
- Deployment configuration table in Prerequisites now covers both A3 and
A2
  - Documents the memory-related differences from the A3 configuration:
    `--max-num-batched-tokens 8192` and `--gpu-memory-utilization 0.92`
(the 95B activated parameters produce large MoE intermediate buffers)
  - Removes FlashComm1 references from the startup commands and the
    optimization table, since the feature has been removed upstream
    (aligned with #13946)

**Stacked on #14570** (docs: refine Qwen3.8-2.4T-A95B tutorial) — this
branch
  is based on that PR's commit and should be merged after it.

  ### Does this PR introduce any user-facing change?

  Yes — documentation only: new A2 deployment instructions in the
  Qwen3.8-2.4T-A95B tutorial.

  ### How was this patch tested?

- The A2 commands were validated on an 8 × Atlas 800 A2 cluster running
the
    W4A8 checkpoint (service startup + functional verification).
  - Markdown lint style checked (no trailing whitespace / tabs; tab-set
    structure mirrors the existing A3 section).

- vLLM version: v0.27.1
- vLLM main:
https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3

---------

Signed-off-by: Erkong <Erkong@noreply.gitcode.com>
Signed-off-by: boes129 <1198231666@qq.com>
Co-authored-by: Erkong <Erkong@noreply.gitcode.com>
B
boes129 committed
b166a6a0ab89a6806a362d5205d31b71fd1fc61e
Parent: bdd3afb
Committed by GitHub <noreply@github.com> on 8/27/2026, 7:20:58 AM