-
Notifications
You must be signed in to change notification settings - Fork 0
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
CUDA FP6 encode/decode goldens (
__nv_fp6_e2m3/e3m2), same pattern as E2M1/E4M3enhancementNew feature or requestNew feature or requestStatus: Open.#32 In SamirMoustafa/torch-floating-point;E0/integer grids: kernel ties-to-even vs Python/JS lower code
bugSomething isn't workingSomething isn't workingStatus: Open.#27 In SamirMoustafa/torch-floating-point;Document tensor-wide pow2 / configurable exponent bias (Round or block=axis; no new encode)
documentationImprovements or additions to documentationImprovements or additions to documentationStatus: Open.#24 In SamirMoustafa/torch-floating-point;Tapered codebook (HiF8 exponent-dependent mantissa; not BlockFormat)
enhancementNew feature or requestNew feature or requestStatus: Open.#23 In SamirMoustafa/torch-floating-point;Document GPTQ/AWQ group-INT4 k=128 (caller block_size; no kernel)
documentationImprovements or additions to documentationImprovements or additions to documentationStatus: Open.#20 In SamirMoustafa/torch-floating-point;Document unsigned E5M3 scale codebook (nearest already; no new encode)
documentationImprovements or additions to documentationImprovements or additions to documentationStatus: Open.#19 In SamirMoustafa/torch-floating-point;Post-block compute-dtype round (reconstruct then FP16/BF16)
enhancementNew feature or requestNew feature or requestStatus: Open.#18 In SamirMoustafa/torch-floating-point;Nested / hierarchical block scales (pair microexponents, not tensor s_global)
enhancementNew feature or requestNew feature or requestStatus: Open.#17 In SamirMoustafa/torch-floating-point;