Full detail view of an issue including body and comments.

Compact one-line summary for list view, with date and key metadata.

Forward pass to dequantize the parameter. Args: quantized_param (`torch.Tensor`): The quantized parameter tensor (from .

Get formatted CUDA/ROCm versions from existing library files.

Get CUDA/HIP version as a string.

Returns the installed version of Gaudi SW.

Load CUDA library XOR CPU, as the latter contains a subset of symbols of the former.

Get ROCm GPU architecture.

Get the path to the XPU native library matching the oneAPI toolchain. Prefers the versioned library (e.g. libbitsandbytes_xpu2026) matching the f

Execute a GraphQL query via the gh CLI, passing the full payload as JSON on stdin.

Create a random batch of input_ids and labels.

Run a few training steps and return peak GPU memory in bytes.

Vectorized 4-bit dequantization: process 8 packed bytes = 16 output values Each byte contains two 4-bit values: high nibble first, low nibble second

ARM NEON FP4 lookup table

ARM NEON NF4 lookup table (16 float values indexed by 4-bit code)

NEON-optimized norm_to_lut_index for 4 float values at a time Maps [-1, 1] → [0, 65535]

Convert a normalized value in [-1, 1] to LUT index [0, 65535]

Pack a dictionary into a torch tensor for storing quant_state items in state_dict. Parameters: - source_dict: The dictionary to be packe

Convert a raw version code string (e.g. '118', '713') to a dotted version (e.g. '11.8', '7.13').

Compare paged_adamw vs adamw numerically.

Compare bnb AdamW vs torch AdamW on CPU to verify correctness.

Train with one optimizer and report results.

Train using HuggingFace Trainer with a bnb optimizer.

Train using HuggingFace Trainer with a bnb optimizer on CPU.

Simple utility to print out environment information.