docs: add validated Jetson Thor DFlash2 serving setup

[marcodelpin]

Summary

• document a tested Qwen3.8-27B DFlash 2 configuration for the 128 GB Jetson AGX Thor Developer Kit
• pin the validated vLLM, NVIDIA PyTorch, FlashInfer, CUDA architecture, weight, and KV-cache formats
• show explicit KV allocation for a 262,144-token context limit
• clarify how --kv-cache-memory-bytes and --gpu-memory-utilization interact in the tested vLLM revision

Validation

The upstream package workflow was reproduced locally on Python 3.12.13:

• built the sdist and wheel
• passed twine check --strict
• installed the generated wheel
• imported dflash
• smoke-tested dflash, dflash generate, and dflash benchmark

The documented configuration was also checked on a live 128 GB Jetson AGX Thor:

• vLLM V2 model runner active
• DFlash2Qwen3ForCausalLM draft model loaded
• text generation and health check successful
• FP8 E4M3 KV cache on sm_110
• max_model_len=262144
• 453,669 tokens of KV capacity, reported as 1.73 concurrent sequences at the 262,144-token limit

The 256K statement describes allocated KV capacity; it does not claim that a 256K-token prompt was executed.

AI assistance disclosure

This documentation change was AI-assisted and has not yet received independent human code review. The commands, package checks, runtime configuration, and reported capacity were verified against the tested hardware and final diff.

[marcodelpin]

Hi - the "Package checks" workflow on this PR is sitting at action_required, so CI hasn't run yet. Could a maintainer approve the run when you get a chance?

It's a docs-only change (README.md), so the wheel build and CLI smoke tests should be unaffected - happy to rebase on current main first if you'd prefer.