feat: support GLM-5.3-Flash NVFP4

[earlvanze]

Adds single-GPU hybrid-offload support for LibertAIDAI/GLM-5.3-Flash-NVFP4.

Highlights:

• GLM-5.3-Flash model/config, weight mapping, tool/reasoning parser compatibility, and benchmark support
• NVFP4 expert-bank loader avoids pathological PyTorch intra-op fan-out while restoring the caller thread count on success or failure
• benchmark flags for MoE CPU thread count and KV reserve sizing

Live validation:

• 1x RTX 5090 32 GB, 2 TiB host RAM, dual AMD EPYC 7V12 on Vast (Italy)
• GPU-local NUMA affinity, 48 MoE CPU threads, hybrid backend, CUDA graph, memory ratio 0.95
• 16.84 steady decode tok/s over 126 measured steps; 59.38 ms/token; 29.21 GiB server VRAM; 51 C peak
• longer 254-step validation: 14.73 tok/s; 67.89 ms/token

Checks:

• uv run pytest tests/models/test_glm5_next_config.py tests/models/test_nvfp4_banks.py -q (5 passed)
• focused Ruff F/I checks passed
• Python compile checks passed

[earlvanze]

Live checkpoint validation found and fixed a KDA prefill layout bug in c61ef42.

causal_conv1d_varlen returns channel-major contiguous storage. The previous transpose left token-major Q/K/V views with a non-unit feature stride, while the Triton KDA kernel indexes the key/head dimensions as contiguous. Materializing the transposed result restores parity.

Actual GLM-5.3-Flash-NVFP4 layer-0 KDA comparison against the Transformers recurrent reference:

• before: cosine 0.0076478, max abs diff 0.0086498
• after: cosine 0.9999971, max abs diff 1.53e-05

Regression checks: 7 passed; Ruff and git diff --check clean. A live RTX PRO 6000 WS / FreeToken offload run now produces the exact requested final answer (PARIS) and completed c8/c16 demand tests with zero failures.

[earlvanze]

Live Vast Serverless qualification completed on commit 3d5354c.

• CA/EU fixed-port template with port 3000 published and Vast model-load estimate set to 200 GiB
• Fresh worker cold load: 262.69 seconds
• Cached activation selected FREETOKEN_PROVISION_STAGE=fast_resume and reported ready in about 109 seconds, inside Vasts fixed 300-second starting window
• Semantic request: HTTP 200 in 13.761 seconds, finish_reason=stop, exact visible response GLM53_SERVERLESS_FAST_RESUME_OK
• Vast four-request worker benchmark: 20.572 workload/s
• Targeted Vast Serverless tests: 5 passed; shell syntax and diff checks passed

The worker and endpoint were retained; the inactivity controller is being allowed to park the GPU automatically.