[Feature] Add DeepSeek V3.2 W8A8 INT8 model support

[Lidang-Jiang]

Summary

• Support serving DeepSeek-V3.2-W8A8-INT8-Dynamic on the vLLM 0.19 Kunlun path.
• Add compatibility/fallback paths for dense FlashMLA, MLA prefill/decode, and compressed-tensors W8A8 dynamic INT8 matmul/quantization.
• Add regression coverage for the adapted MLA and quantization paths.

This PR is based on the v0.19 upgrade branch from #315; the new commit adds the DeepSeek V3.2 W8A8 INT8 Dynamic serving adaptation.

Full sanitized logs: https://gist.github.com/Lidang-Jiang/14df1d68a4f9cd17f18a63950cd007ad

Before

Baseline commit: 864d569

Service startup failed before the adaptation:

(APIServer pid=165706)   Value error, This model does not support `--runner generate`. [type=value_error, input_value=ArgsKwargs((), {'model': ...nderer_num_workers': 1}), input_type=ArgsKwargs]

Client verification:

mode=before
code=864d569
service_log=/workspace/deepseek_v32_before_864d569.log
curl /v1/models:
curl: (7) Failed to connect to 127.0.0.1 port 8566: Connection refused

curl /v1/chat/completions:
curl: (7) Failed to connect to 127.0.0.1 port 8566: Connection refused

After

Service startup:

(Worker_TP0 pid=166770) INFO 04-24 20:47:36 [default_loader.py:384] Loading weights took 106.27 seconds
(Worker_TP0 pid=166770) INFO 04-24 20:47:37 [gpu_model_runner.py:4820] Model loading took 81.07 GiB memory and 107.374742 seconds
(EngineCore pid=166479) INFO 04-24 20:48:46 [kv_cache_utils.py:1319] GPU KV cache size: 44,352 tokens
(EngineCore pid=166479) INFO 04-24 20:48:46 [kv_cache_utils.py:1324] Maximum concurrency for 32,768 tokens per request: 1.35x
(APIServer pid=166227) INFO:     Application startup complete.

Client verification:

mode=after
commit=3dffb27
service_log=/workspace/deepseek_v32_w8a8_p8566.log
curl /health:
200
curl /v1/models:
{"object":"list","data":[{"id":"DeepSeek-V3.2-W8A8-INT8-Dynamic","object":"model","owned_by":"vllm","root":"/models/DeepSeek-V3.2-W8A8-INT8-Dynamic","max_model_len":32768}]}

curl /v1/chat/completions:
HTTP 200, assistant content: 我是DeepSeek，由深度求索公司创造的AI助手！

Unit tests:

============================= test session starts ==============================
platform linux -- Python 3.10.15, pytest-8.3.3, pluggy-1.5.0
rootdir: /workspace/vLLM-Kunlun
configfile: pyproject.toml
plugins: typeguard-4.4.1, repeat-0.9.3, anyio-4.6.2.post1, timeout-2.3.1, metadata-3.1.1, html-3.2.0, cov-7.0.0
collected 59 items

tests/ut/test.py ....................................................... [ 93%]
....                                                                     [100%]

============================== 59 passed in 6.91s ==============================

Test plan

• /workspace/python310_torch25_cuda/bin/python -m pytest tests/ut/test.py
• Start service with DeepSeek-V3.2-W8A8-INT8-Dynamic
• curl http://127.0.0.1:8566/health
• curl http://127.0.0.1:8566/v1/models
• curl http://127.0.0.1:8566/v1/chat/completions