Spaces:
Running on Zero
Running on Zero
Commit History
-
Avoid fp32 logits copy in ZeroGPU GRPO loss Running
Verified signature63cf098 -
Patch ZeroGPU loss logprob memory path
Verified signature724e766 -
Update ZeroGPU GRPO prototype with memory-safe logprobs
Verified signature8078b9c -
Pin flash-linear-attention for ZeroGPU smoke
Verified signature355e365 -
Adapt FLA patch to installed kernel signature
Verified signature31a73cc -
Repair incomplete ScatterMoE LoRA kernel metadata
Verified signatured7bf13c -
Patch ZeroGPU kernel metadata recovery
Verified signaturee29a7af -
Use kernels 0.14 for Kernel Hub ScatterMoE LoRA
Verified signature53b9a59 -
Update Hatchery ZeroGPU prototype for ScatterMoE MoE smoke
Verified signaturebd2054b
-
Try causal-conv1d 1.6.2 post1
Verified signature970532d -
Restore no-CCE FLA wrapper with transformers pin
Verified signatureffe6a1e -
Pin transformers and remove CCE
Verified signature72f8c36 -
Pass FLA grouped constexpr args
Verified signatureaa49223 -
Fix FLA grouped value-head backward patch
Verified signature1038952 -
Make FLA backward patch contiguous
Verified signature492d6f1 -
Install cut-cross-entropy and log CUDA memory
Verified signature9ff4e9a
-
unwrap fla autotuner to raw jit kernel
Verified signaturec5d7d6f -
patch fla bwd fixed kernel launch
Verified signature3501891 -
bypass unsafe fla triton autotune on zerogpu
Verified signature7de01ec -
shim glibc mathcalls for tilelang nvcc
Verified signature49cc0c2 -
narrow tilelang nvcc c2x workaround
Verified signature8742f0e -
patch tilelang tvm nvcc compile calls
Verified signature30230d2 -
patch tilelang nvcc glibc c2x flags
Verified signature2c3e189 -
patch cuda math function redeclarations
Verified signature0f804f0 -
return zerogpu gpu worker tracebacks
Verified signaturef5f7a6e -
return zerogpu train debug errors
Verified signaturefd91bef -
add zerogpu train debug endpoint
Verified signatureb67c605 -
surface zerogpu endpoint errors
Verified signature160ab3a -
log zerogpu endpoint tracebacks
Verified signatured7560d4 -
use writable cuda header shim for tilelang nvcc
Verified signature60910df -
patch cuda math header for tilelang nvcc
Verified signature8ea51e2 -
use public scattermoe kernel ref
Verified signatured3caeb4 -
enable core scattermoe tilelang dflash fixes
Verified signature715a73d -
add diagnostic tensor prefix breakdown
Verified signaturea10bbc7 -
expose model diagnostics in gradio api
Verified signatureefc4392 -
add model diagnostics endpoint
Verified signature065340b -
Use batched dflash fork
Verified signaturedb2e429
-
Expose training profiles for wandb validation
Verified signature97a96c6 -
Pin hatchery core dflash metadata commit
Verified signature723a7bc -
Manually place TorchAO FP8 tensors for ZeroGPU
Verified signature878779c -
Move TorchAO FP8 base before LoRA wrapping
Verified signature5f5a708 -
Patch TorchAO Float8Tensor conversion ops for ZeroGPU
Verified signatureda0b39f -
Patch TorchAO Float8Tensor empty_like for ZeroGPU
Verified signature5f78613 -
Deploy TorchAO FP8 ZeroGPU prototype
Verified signaturefa47cb8
-
Use stable ZeroGPU state and sample logprobs
Verified signature28dc00c -
Return rollout logprobs from ZeroGPU sample
Verified signature958e819 -
Use kernels-provided FA2 dependency
Verified signaturef34e1e8 -
Enable FA2 ZeroGPU prototype service
Verified signaturec1c9b79 -
Enable FA2 ZeroGPU prototype service
Verified signature9cee980 -
Install causal conv wheel and skip zero-advantage GRPO groups
Verified signaturea6ad39a