[Bugfix][SM70] Enable Qwen3.8 AWQ in dual-compile lane

[Leonccaa]

Purpose

Follow up on #477 by admitting the Qwen3.8 AWQ text-only runtime to the
existing SM70 dual-compile lane.

The checkpoint keeps the outer Qwen4ExpForConditionalGeneration
architecture even when it is loaded with --language-model-only, while the
#477 contract admitted only Qwen4ExpForCausalLM. This change also accepts the
conditional-generation wrapper, but only when
multimodal_config.language_model_only is explicitly true. Full multimodal
loading and a missing multimodal config remain rejected.

Real TP4 validation then exposed a second late-binding issue. #477 enables
hybrid PLE from VllmConfig.__post_init__, after ParallelConfig validation
would normally allocate the shared PLE IPC endpoint. As a result, all workers
attempted to connect to an empty address and failed with
zmq.error.ZMQError: Invalid argument (addr='').

This PR makes PLE endpoint initialization idempotent and invokes it when the
model-aware defaults enable PLE late. Explicit PLE configuration retains the
same behavior.

There are no AWQ kernel, quantization-math, weight-format, or scheduler
changes. Operator capability checks and existing fallbacks remain in place.
In particular, this PR does not address mixed-phase latency from chunked
prefill scheduling.

Test Plan

• Run the focused dual-compile/config tests on the final branch.
• Run changed-file pre-commit checks.
• Start the real Qwen3.8 native-g32 AWQ checkpoint on 4x V100, TP4, FP16
  activations, KV auto, MTP0, prefix caching off, and
  FULL_AND_PIECEWISE compilation without explicitly setting the PLE envs.
• Exercise C1 x 64K, C4 x 64K, and C8 x 16K once per cell with streaming token
  timestamps, 320 requested output tokens, and ignore_eos=false.
• Repeat the same runtime contract with [Perf][SM70] Add indexed-A AWQ prefill for Qwen3.8 #464/perf(sm70): right-size AWQ MoE persistent scratch #465/[Perf][SM70] Bound Qwen3.8 AWQ W2 output scratch #470/[Perf][SM70] Compact Qwen3.8 AWQ scale metadata #473 enabled together
  as an integration check.

Test Result

• .venv/bin/pytest -q --confcutdir=tests/compile tests/compile/test_sm70_decode_graph.py:
  7 passed.
• Changed-file pre-commit suite: all applicable hooks passed, including Ruff,
  formatting, mypy, SPDX, configuration validation, and forbidden-import
  checks.
• Before the IPC fix, the real model completed weight loading but all four GPU
  workers reproduced the empty-address ZMQ failure.
• After the fix, both the post-[Core][SM70] Split Qwen3.8 prefill and decode compilation #477 main-only arm and the four-PR integration
  arm completed startup. All four workers used the same non-empty IPC endpoint,
  PLE registered 4/4 workers, and the service built the large prefill graph plus
  decode capture sizes [1, 8].
• Each arm completed all 13 requests in the three-cell contract. Every request
  returned 320 tokens with finish_reason=length; no traceback, OOM, or
  runtime worker exit was observed.
• The four-PR integration arm retained its memory benefit: model loading
  21.24 -> 20.24 GiB/rank and KV capacity 386,392 -> 505,574 tokens
  (+30.85%) versus the matched post-[Core][SM70] Split Qwen3.8 prefill and decode compilation #477 main-only arm.
• The same run still measured C4 x 64K mixed-phase mean ITL at 2.841 s on
  main-only and 2.719 s on the four-PR stack. That scheduler behavior remains
  separate follow-up work.

AI assistance disclosure: OpenAI Codex assisted with implementation, test
execution, runtime evidence collation, and PR preparation. The submitter
reviewed the changes and validation results.

---

Essential Elements of an Effective PR Description Checklist

• The purpose and relationship to [Core][SM70] Split Qwen3.8 prefill and decode compilation #477 are documented.
• The test plan and commands are documented.
• Unit, pre-commit, and real V100 results are documented.
• Scope and remaining scheduler boundary are documented.

[yangzhuxinyzx]

已基于包含 #492 的最新 main ba7f6ff 审计和模拟合并。确认此次改动只扩展已有双编译的 LM-only wrapper 准入并补齐 PLE IPC 初始化，不改变 AWQ 运算或采样。7 项定向测试通过，修改文件全部 pre-commit 通过；自动启用及重复初始化共享同一个非空 IPC 地址。结合 PR 的 TP4 启动证据，审计通过。