[feat] Support MetaX hardware backend

[ILikeIneine]

PR Category

PR Types

PR Description

Support Qwen3-next on MetaX hardware backend

[CLAassistant]

All committers have signed the CLA.

[ILikeIneine]

qwen3-next-80b-a3b

Currently the models on MetaX hardware backend can only work with:

• vllm's original gpu worker vllm.v1.worker.gpu_worker.Worker (not WorkerFL)
• eager mode (with --enforce-eager args)
• disable flag-gems

USE_FLAGGEMS=0 vllm serve ${qwen3-next-model_path} --enforce-eager

Verified by @ceci3

Todos

Here are the remaining issues on supporting metax backend.

1. Big mess on config's name, vendor name, platform name, backend name, backend vendor, impl_id (needs a more clarified relationship)
2. Can't enable CudaGraph on metax backend (CudaGraph.MODE is always None)
3. Metax backend needs an entry for its patches (The patches need to be applied and take effect on every subprocess that forks or respawned by vllm, like gpu workers and engine core)
4. gdn_attn has cuda_graph bug on --enforce-eager mode. Need a hotfix for the file.

Tested Scenario

model	eager mode (--enforce-eager)	worker	flag_gems(4.2.0)	result	Note
qwen3-0.6b	OFF	FL's worker	enable	✅	worked but no cuda graph capture (capture list is empty)
qwen3-next	ON	vllm's origin worker	disable	✅	worked but no cuda graph capture (capture list is empty)
qwen3-next	OFF	vllm's origin worker	disable	❌	crash with torch compile backends due to cudagraph not currectly enabled
qwen3-next	ON	FL's worker	disable	❌	crash without any informations
qwen3-next	ON	vllm's origin worker	enable	❌	torch crashed
qwen3-next	ON	vllm's origin worker	enable (with 4.2.1rc0 latest	❌	confilct on register _C custom_ops (silu_and_mul)