fix(quantizer): make _pack/_unpack_qjl_signs CUDA-graph-safe

[AEON-7]

Summary

Fix an unpinned CPU → GPU copy in the quantizer hot path that breaks
CUDA graph capture. Both capture_only and hybrid modes currently
crash during vLLM's determine_available_memory() unless the server
is launched with --enforce-eager (which costs 20–30% decode throughput).

Root cause

turboquant/quantizer.py:216-224 —
_pack_qjl_signs and _unpack_qjl_signs allocate the powers-of-two
lookup tensor inside the function:

powers = torch.tensor([1, 2, 4, 8, 16, 32, 64, 128],
                       device=signs.device, dtype=torch.uint8)

torch.tensor([...], device=cuda) creates the tensor on CPU first then
copies to GPU. That's an unpinned host-to-device copy, which PyTorch
refuses inside a CUDA graph capture region:

RuntimeError: Cannot copy between CPU and CUDA tensors during CUDA
graph capture unless the CPU tensor is pinned. Please use
tensor.pin_memory() or allocate the tensor with pin_memory=True.

Repro

# any vllm serve with TQ installed via collective_rpc on the worker:
vllm serve <model> --attention-backend flash_attn \
    --gpu-memory-utilization 0.85 --max-num-seqs 16
# then: executor.collective_rpc(lambda w: install_turboquant_hooks(
#           w.model_runner, mode="capture_only"))

Without this patch the engine crashes during
_warmup_and_capture → _dummy_run. With --enforce-eager, the plugin
works but at a substantial perf cost.

Fix

Cache the powers tensor per-device in a module-level dict. Allocate
once, reuse forever. No hot-path host-device copy.

Validation

Tested on NVIDIA DGX Spark (GB10 / sm_121a) with
ghcr.io/aeon-7/vllm-dflash:latest, Qwen3.5-27B NVFP4 hybrid model
(16 full-attention + 48 linear-attention layers), vLLM 0.19.1rc1.
Both modes now boot cleanly with CUDA graphs enabled.

Throughput vs baseline (no TQ) at full production config

Natural-language prompts, deterministic (temperature=0), steady-state
post-warmup, 16 requests per level.

Concurrency	TQ off (ref)	TQ capture_only	TQ hybrid
c=1 (code)	64.02 tok/s	61.50 tok/s	61.71 tok/s
c=4 (code)	181.47 tok/s	175.71 tok/s	175.79 tok/s
c=8 (code)	262.77 tok/s	255.19 tok/s	252.78 tok/s
c=16 (code)	327.89 tok/s	314.93 tok/s	318.36 tok/s
c=1 (prose)	29.46 tok/s	28.14 tok/s	28.49 tok/s
c=16 (prose)	151.81 tok/s	147.43 tok/s	148.80 tok/s

Measured overhead: ~3% across all modes, concurrencies, and prompt
styles — matching the advertised profile.

Test plan for reviewer

• pytest test_modular.py test_turboquant.py validate_paper.py still passes on CUDA
• Boot a vLLM server with install_turboquant_hooks(..., mode="hybrid") without --enforce-eager; engine should reach Application startup complete without the "Cannot copy between CPU and CUDA tensors" error
• No regression on the existing benchmark.py / proof.py flows

🤖 Generated with Claude Code