Module kv_dequant 

Host-side dequantization of paged KV cache blocks → BF16.

Used by --high-speed-swap Phase 6.2.c to produce BF16 source data for the orchestrator’s tile-streaming attention kernel from any of Atlas’s quantized KV layouts. The kernel-side packing/scale layouts mirrored here:

Quant	data bytes/elem	scale	LUT	source kernel
BF16	2	none	identity	(direct copy, not in this module)
FP8	1 (E4M3)	tensor	e4m3_lut	reshape_and_cache_fp8.cu
NVFP4	0.5 (4-bit)	group	NVFP4_E2M1_LUT	paged_decode_attn_nvfp4.cu
Turbo4	0.5 (4-bit)	group	TURBO4_LUT (Lloyd-Max 16)	paged_decode_attn_turbo4.cu
Turbo3	0.375 (3-bit)	group	TURBO3_LUT (Lloyd-Max 8)	paged_decode_attn_turbo3.cu
Turbo8	1 (E4M3)	group	e4m3_lut	paged_decode_attn_turbo8.cu

“Group” scales cover 16 elements per scale byte (NVFP4_GROUP_SIZE) and are stored in a separate section after the data section within each block. All LUTs match their CUDA-side counterparts byte-for-byte.

Constants

NVFP4_E2M1_LUT

E2M1 4-bit codebook (NVFP4). Matches kernels/gb10/common/paged_decode_attn_nvfp4.cu:118.

NVFP4_GROUP_SIZE

Group size for per-group FP8 scales. Matches NVFP4_GROUP_SIZE in the per-quant attention kernels.

TURBO3_LUT

Turbo3 8-level Lloyd-Max codebook. Matches kernels/gb10/common/paged_decode_attn_turbo3.cu:137.

TURBO4_LUT

Turbo4 16-level Lloyd-Max codebook. Matches kernels/gb10/common/paged_decode_attn_turbo4.cu:121.

Functions

dequant_4bit_block_to_bf16

Dequant a 4-bit packed (NVFP4 or Turbo4) KV block to BF16.

dequant_fp8_to_bf16

Dequant FP8 (E4M3) bytes to BF16, applying a per-tensor scale.

dequant_turbo3_block_to_bf16

Dequant a Turbo3 (3-bit packed) KV block to BF16.

dequant_turbo8_block_to_bf16

Dequant a Turbo8 (FP8 E4M3 data + per-group BF16 scales) KV block to BF16.

e4m3_lut

Borrow the compile-time table.