Avoid fp32 logits copy in ZeroGPU GRPO loss

Patch ZeroGPU loss logprob memory path

Update ZeroGPU GRPO prototype with memory-safe logprobs

Pin flash-linear-attention for ZeroGPU smoke

Adapt FLA patch to installed kernel signature

Repair incomplete ScatterMoE LoRA kernel metadata

Patch ZeroGPU kernel metadata recovery

Use kernels 0.14 for Kernel Hub ScatterMoE LoRA

Update Hatchery ZeroGPU prototype for ScatterMoE MoE smoke

Try causal-conv1d 1.6.2 post1

Restore no-CCE FLA wrapper with transformers pin

Pin transformers and remove CCE

Pass FLA grouped constexpr args

Fix FLA grouped value-head backward patch

Make FLA backward patch contiguous

Install cut-cross-entropy and log CUDA memory

unwrap fla autotuner to raw jit kernel

patch fla bwd fixed kernel launch

bypass unsafe fla triton autotune on zerogpu

shim glibc mathcalls for tilelang nvcc

narrow tilelang nvcc c2x workaround

patch tilelang tvm nvcc compile calls

patch tilelang nvcc glibc c2x flags

patch cuda math function redeclarations

return zerogpu gpu worker tracebacks

return zerogpu train debug errors

add zerogpu train debug endpoint

surface zerogpu endpoint errors

log zerogpu endpoint tracebacks

use writable cuda header shim for tilelang nvcc

patch cuda math header for tilelang nvcc

use public scattermoe kernel ref

enable core scattermoe tilelang dflash fixes

add diagnostic tensor prefix breakdown

expose model diagnostics in gradio api

add model diagnostics endpoint

Use batched dflash fork

Expose training profiles for wandb validation

Pin hatchery core dflash metadata commit

Manually place TorchAO FP8 tensors for ZeroGPU

Move TorchAO FP8 base before LoRA wrapping

Patch TorchAO Float8Tensor conversion ops for ZeroGPU

Patch TorchAO Float8Tensor empty_like for ZeroGPU

Deploy TorchAO FP8 ZeroGPU prototype

Use stable ZeroGPU state and sample logprobs

Return rollout logprobs from ZeroGPU sample

Use kernels-provided FA2 dependency

Enable FA2 ZeroGPU prototype service

Enable FA2 ZeroGPU prototype service

Install causal conv wheel and skip zero-advantage GRPO groups