[vllm_fp8_utils] fix: preserve vLLM>=0.12 FP8 Marlin workspace init

[JohnConnor123]

Summary

• Fix vLLM FP8 (Marlin) crash in verl: missing workspace attribute (seen as AttributeError during engine init / profiling).
• Root cause: verl monkey-patched vLLM FP8 process_weights_after_loading and (for vLLM>=0.12) bypassed vLLM-native Marlin preparation, where vLLM allocates/sets layer.workspace.
• Fix: for vLLM>=0.12 wrap and call the original vLLM post-load method (preserving Marlin workspace init) and restore verl-specific custom Parameter attrs; additionally ensure QKVParallelLinear instances always have a workspace attribute (lazy + idempotent) to avoid early AttributeError.

Reproduction

The bug shows up on the Marlin FP8 path (typically GPUs without native FP8 support). If you are on a native-FP8 GPU, you can force Marlin for testing (see below).

Minimal repro script (crashes on main, fixed on this PR)

set -euo pipefail

# Force Marlin even on native-FP8 GPUs (H100/SM90+) so the bug reproduces deterministically.
export VLLM_TEST_FORCE_FP8_MARLIN=1

python3 - <<'PY'
from verl.utils.vllm.vllm_fp8_utils import apply_vllm_fp8_patches
apply_vllm_fp8_patches()

from vllm import LLM, SamplingParams

llm = LLM(
    model="Qwen/Qwen3-1.7B",
    tensor_parallel_size=1,
    dtype="bfloat16",
    quantization="fp8",
    trust_remote_code=True,
)
out = llm.generate(["Hello"], SamplingParams(max_tokens=8))
print(out[0].outputs[0].text)
PY

Expected error (before)

AttributeError: 'QKVParallelLinear' object has no attribute 'workspace'

(Triggered from the Marlin FP8 execution path where vLLM uses workspace=layer.workspace; see vLLM v0.12.0 Fp8LinearMethod.apply:
https://raw.githubusercontent.com/vllm-project/vllm/v0.12.0/vllm/model_executor/layers/quantization/fp8.py.)

Notes on GPU support (native FP8 vs Marlin)

• On GPUs without native FP8, vLLM commonly falls back to Marlin weight-only FP8, which requires a workspace buffer.
• On GPUs with native FP8, vLLM usually uses a non-Marlin FP8 path and does not rely on layer.workspace, so this crash often does not appear.
• You can force the Marlin path for testing on native-FP8 GPUs with VLLM_TEST_FORCE_FP8_MARLIN=1 (as in the repro script above).

Test plan

• Run the script above on main → observe the workspace AttributeError (or force Marlin).
• Run the same script on this PR branch → it should complete and print a short generation.

[CLAassistant]

All committers have signed the CLA.

[wuxibin89]

@JohnConnor123 Can you clarify what GPU you're using? I have reservations about supporting fp8 rollout on GPUs without native fp8 capability.

[JohnConnor123]

@wuxibin89 I use 2 rtx3090.