Fix MXFP4 triton kernel launch on triton>=3.x (seed=None -> int)

[kylefoxaustin]

Problem

Pseudo-quantization training with QuestMXFP4Quantizer (the config in main_setup.sh) crashes at the first forward pass on recent triton:

TypeError: 'NoneType' object cannot be interpreted as an integer
  .../triton/backends/nvidia/driver.py", line 713, in __call__

Root cause

In mxfp4_forward_kernel_wrapper, seed is set to None when stochastic_round is False:

if stochastic_round:
    seed = randint(0, 1000000)
else:
    seed = None          # passed to the kernel's `seed: int` (non-constexpr) arg

The kernel declares seed: int (non-constexpr), so triton tries to pack it into the launch arguments. Older triton tolerated an unused None; triton >= 3.x rejects it at launch time. Since seed is only read inside the if stochastic_round: branch (dead-code-eliminated when False), the value is never used in this path.

Fix

Pass 0 instead of None — behavior-preserving (the value is unread when stochastic_round=False).

Testing

With this one-line change, the full recipe (QuestMXFP4 weights+activations + AlbertTsengQuantizer 4-bit gradients + the Q(E)Q(Wt)t_Q(Et)Q(Xt)t backward scheme) trains end-to-end — verified on triton 3.6 / torch 2.11+cu128 / RTX 5090 (sm_120), 30M Llama on WikiText-103, FP4 val-loss tracking the BF16 baseline to within ~0.07 nats.