SD15 MACaQ Pruned Full-Width

72.84% smaller U-Net weight payload ยท 1,719.04 MB FP16 โ†’ 466.96 MB (445.32 MiB) packed.

This is an MXFP4 W4A4 quantization of Stable Yogi SD 1.5 V9 LCM. It retains the original U-Net layer dimensions. The repository name follows the earlier MACaQ release. The storage savings come from weight quantization, and all U-Net layers remain present.

Weights

Component Precision Tensor payload
U-Net MXFP4 W4A4 with retained higher-precision tensors 466.96 MB
CLIP text encoder FP16 246.12 MB
VAE FP16 167.31 MB
Total weight files including headers 880.62 MB

Full-width calibration

The U-Net was refit from the original FP16 checkpoint. Calibration sampled activations from 32 prompts with up to 128 rows per forward call. Each of the 184 linear layers and 98 convolutions used a Hessian that spans its full input width, including flattened convolution patch channels, with alignment padding to multiples of 32 where needed. The widest layer used 23,040 inputs. The earlier release used 256-wide convolution Hessian blocks.

The fit balances input channels, applies a blockwise Hadamard rotation, screens block exponents, then refines adjacent FP4 weight codes using the full Hessian at fixed exponents. Each row is refined to a local stopping criterion. Per-layer sample counts, widths, flip counts, and objectives are in the run report.

FP16 comparisons

Six matched prompt-and-seed pairs were rendered at 512ร—512 using LCM, 4 steps, and CFG 1.2. The left image is FP16 and the right image uses the quantized U-Net. Both use the same CLIP text encoder and VAE. The quantized column used emulated MXFP4 W4A4 on the L4, with FP4 rounding and BF16 matrix multiplication.

The mean LPIPS to FP16 was 0.4292 across these six pairs. Lower values indicate closer image appearance. Some details change, including glasses in the train-station scene and the car's viewing angle. The earlier release's results were reported with SSIM, so this LPIPS result does not rank the two releases.

Seed 7001 ยท Candid photo of an adult man holding a newspaper at a train station ยท LPIPS 0.3830

FP16 and MXFP4 images for seed 7001

Seed 7002 ยท A full-body photograph of an adult woman in a long blue coat at dusk ยท LPIPS 0.3491

FP16 and MXFP4 images for seed 7002

Seed 7003 ยท Close-up of a ceramic teapot, two cups, and a linen napkin on a table ยท LPIPS 0.3947

FP16 and MXFP4 images for seed 7003

Seed 7004 ยท A red vintage car parked on a curve in a desert highway, sunny afternoon ยท LPIPS 0.4756

FP16 and MXFP4 images for seed 7004

Seed 7005 ยท Portrait of an adult architect wearing black rectangular glasses in a sunlit studio ยท LPIPS 0.4408

FP16 and MXFP4 images for seed 7005

Seed 7006 ยท Exactly three green apples beside one yellow pear on a wooden board ยท LPIPS 0.5322

FP16 and MXFP4 images for seed 7006

U-Net layer inventory

The JSON inventory includes every stored U-Net tensor.

Layer group Modules Storage (MiB)
3ร—3 convolutions 52 276.21
Normalization 109 0.38
1ร—1 convolutions 46 25.66
Attention projections 128 50.06
MLP projections 32 78.63
Timestep and other linears 24 14.38
3ร—3 convolutions ยท 52 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
conv_in MXFP4 W4A4 320ร—4ร—3ร—3 64 46,720 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—128], W_scales U8 [320ร—8], bias BF16 [320], pre_scale BF16 [256]
conv_out MXFP4 W4A4 4ร—320ร—3ร—3 2,880 14,728 B_inv_T BF16 [32ร—32], W_packed U8 [4ร—1536], W_scales U8 [4ร—96], bias BF16 [4], pre_scale BF16 [3072]
down_blocks.0.downsamplers.0.conv MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
down_blocks.0.resnets.0.conv1 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
down_blocks.0.resnets.0.conv2 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
down_blocks.0.resnets.1.conv1 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
down_blocks.0.resnets.1.conv2 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
down_blocks.1.downsamplers.0.conv MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
down_blocks.1.resnets.0.conv1 MXFP4 W4A4 640ร—320ร—3ร—3 2,880 1,053,952 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1536], W_scales U8 [640ร—96], bias BF16 [640], pre_scale BF16 [3072]
down_blocks.1.resnets.0.conv2 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
down_blocks.1.resnets.1.conv1 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
down_blocks.1.resnets.1.conv2 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
down_blocks.2.downsamplers.0.conv MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.2.resnets.0.conv1 MXFP4 W4A4 1280ร—640ร—3ร—3 5,760 4,020,224 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2944], W_scales U8 [1280ร—184], bias BF16 [1280], pre_scale BF16 [5888]
down_blocks.2.resnets.0.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.2.resnets.1.conv1 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.2.resnets.1.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.3.resnets.0.conv1 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.3.resnets.0.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.3.resnets.1.conv1 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
down_blocks.3.resnets.1.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
mid_block.resnets.0.conv1 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
mid_block.resnets.0.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
mid_block.resnets.1.conv1 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
mid_block.resnets.1.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.0.resnets.0.conv1 MXFP4 W4A4 1280ร—2560ร—3ร—3 23,040 15,717,888 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—11520], W_scales U8 [1280ร—720], bias BF16 [1280], pre_scale BF16 [23040]
up_blocks.0.resnets.0.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.0.resnets.1.conv1 MXFP4 W4A4 1280ร—2560ร—3ร—3 23,040 15,717,888 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—11520], W_scales U8 [1280ร—720], bias BF16 [1280], pre_scale BF16 [23040]
up_blocks.0.resnets.1.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.0.resnets.2.conv1 MXFP4 W4A4 1280ร—2560ร—3ร—3 23,040 15,717,888 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—11520], W_scales U8 [1280ร—720], bias BF16 [1280], pre_scale BF16 [23040]
up_blocks.0.resnets.2.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.0.upsamplers.0.conv MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.1.resnets.0.conv1 MXFP4 W4A4 1280ร—2560ร—3ร—3 23,040 15,717,888 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—11520], W_scales U8 [1280ร—720], bias BF16 [1280], pre_scale BF16 [23040]
up_blocks.1.resnets.0.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.1.resnets.1.conv1 MXFP4 W4A4 1280ร—2560ร—3ร—3 23,040 15,717,888 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—11520], W_scales U8 [1280ร—720], bias BF16 [1280], pre_scale BF16 [23040]
up_blocks.1.resnets.1.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.1.resnets.2.conv1 MXFP4 W4A4 1280ร—1920ร—3ร—3 17,280 11,876,864 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—8704], W_scales U8 [1280ร—544], bias BF16 [1280], pre_scale BF16 [17408]
up_blocks.1.resnets.2.conv2 MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.1.upsamplers.0.conv MXFP4 W4A4 1280ร—1280ร—3ร—3 11,520 7,861,248 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—5760], W_scales U8 [1280ร—360], bias BF16 [1280], pre_scale BF16 [11520]
up_blocks.2.resnets.0.conv1 MXFP4 W4A4 640ร—1920ร—3ร—3 17,280 5,956,864 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—8704], W_scales U8 [640ร—544], bias BF16 [640], pre_scale BF16 [17408]
up_blocks.2.resnets.0.conv2 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
up_blocks.2.resnets.1.conv1 MXFP4 W4A4 640ร—1280ร—3ร—3 11,520 3,943,168 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—5760], W_scales U8 [640ร—360], bias BF16 [640], pre_scale BF16 [11520]
up_blocks.2.resnets.1.conv2 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
up_blocks.2.resnets.2.conv1 MXFP4 W4A4 640ร—960ร—3ร—3 8,640 2,980,096 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—4352], W_scales U8 [640ร—272], bias BF16 [640], pre_scale BF16 [8704]
up_blocks.2.resnets.2.conv2 MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
up_blocks.2.upsamplers.0.conv MXFP4 W4A4 640ร—640ร—3ร—3 5,760 2,017,024 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—2944], W_scales U8 [640ร—184], bias BF16 [640], pre_scale BF16 [5888]
up_blocks.3.resnets.0.conv1 MXFP4 W4A4 320ร—960ร—3ร—3 8,640 1,499,776 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—4352], W_scales U8 [320ร—272], bias BF16 [320], pre_scale BF16 [8704]
up_blocks.3.resnets.0.conv2 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
up_blocks.3.resnets.1.conv1 MXFP4 W4A4 320ร—640ร—3ร—3 5,760 1,015,424 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—2944], W_scales U8 [320ร—184], bias BF16 [320], pre_scale BF16 [5888]
up_blocks.3.resnets.1.conv2 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
up_blocks.3.resnets.2.conv1 MXFP4 W4A4 320ร—640ร—3ร—3 5,760 1,015,424 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—2944], W_scales U8 [320ร—184], bias BF16 [320], pre_scale BF16 [5888]
up_blocks.3.resnets.2.conv2 MXFP4 W4A4 320ร—320ร—3ร—3 2,880 531,072 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—1536], W_scales U8 [320ร—96], bias BF16 [320], pre_scale BF16 [3072]
Normalization ยท 109 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
conv_norm_out F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.0.norm F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.0.transformer_blocks.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.0.transformer_blocks.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.0.transformer_blocks.0.norm3 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.1.norm F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.1.transformer_blocks.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.1.transformer_blocks.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.attentions.1.transformer_blocks.0.norm3 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.resnets.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.resnets.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.resnets.1.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.0.resnets.1.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.1.attentions.0.norm F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.0.transformer_blocks.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.0.transformer_blocks.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.0.transformer_blocks.0.norm3 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.1.norm F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.1.transformer_blocks.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.1.transformer_blocks.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.attentions.1.transformer_blocks.0.norm3 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.resnets.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
down_blocks.1.resnets.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.resnets.1.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.1.resnets.1.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.2.attentions.0.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.1.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.resnets.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
down_blocks.2.resnets.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.resnets.1.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.2.resnets.1.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.3.resnets.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.3.resnets.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.3.resnets.1.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
down_blocks.3.resnets.1.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.attentions.0.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.attentions.0.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.attentions.0.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.attentions.0.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.resnets.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.resnets.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.resnets.1.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
mid_block.resnets.1.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.0.resnets.0.norm1 F16 2560 โ€” 10,240 bias F16 [2560], weight F16 [2560]
up_blocks.0.resnets.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.0.resnets.1.norm1 F16 2560 โ€” 10,240 bias F16 [2560], weight F16 [2560]
up_blocks.0.resnets.1.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.0.resnets.2.norm1 F16 2560 โ€” 10,240 bias F16 [2560], weight F16 [2560]
up_blocks.0.resnets.2.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.0.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.1.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.2.norm F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.norm3 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.resnets.0.norm1 F16 2560 โ€” 10,240 bias F16 [2560], weight F16 [2560]
up_blocks.1.resnets.0.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.resnets.1.norm1 F16 2560 โ€” 10,240 bias F16 [2560], weight F16 [2560]
up_blocks.1.resnets.1.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.1.resnets.2.norm1 F16 1920 โ€” 7,680 bias F16 [1920], weight F16 [1920]
up_blocks.1.resnets.2.norm2 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.2.attentions.0.norm F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.0.transformer_blocks.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.0.transformer_blocks.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.0.transformer_blocks.0.norm3 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.1.norm F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.1.transformer_blocks.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.1.transformer_blocks.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.1.transformer_blocks.0.norm3 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.2.norm F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.2.transformer_blocks.0.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.2.transformer_blocks.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.attentions.2.transformer_blocks.0.norm3 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.resnets.0.norm1 F16 1920 โ€” 7,680 bias F16 [1920], weight F16 [1920]
up_blocks.2.resnets.0.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.resnets.1.norm1 F16 1280 โ€” 5,120 bias F16 [1280], weight F16 [1280]
up_blocks.2.resnets.1.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.2.resnets.2.norm1 F16 960 โ€” 3,840 bias F16 [960], weight F16 [960]
up_blocks.2.resnets.2.norm2 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.3.attentions.0.norm F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.0.transformer_blocks.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.0.transformer_blocks.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.0.transformer_blocks.0.norm3 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.1.norm F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.1.transformer_blocks.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.1.transformer_blocks.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.1.transformer_blocks.0.norm3 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.2.norm F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.2.transformer_blocks.0.norm1 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.2.transformer_blocks.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.attentions.2.transformer_blocks.0.norm3 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.resnets.0.norm1 F16 960 โ€” 3,840 bias F16 [960], weight F16 [960]
up_blocks.3.resnets.0.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.resnets.1.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.3.resnets.1.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
up_blocks.3.resnets.2.norm1 F16 640 โ€” 2,560 bias F16 [640], weight F16 [640]
up_blocks.3.resnets.2.norm2 F16 320 โ€” 1,280 bias F16 [320], weight F16 [320]
1ร—1 convolutions ยท 46 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
down_blocks.0.attentions.0.proj_in MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.0.proj_out MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.1.proj_in MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.1.proj_out MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.1.attentions.0.proj_in MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.0.proj_out MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.1.proj_in MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.1.proj_out MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.resnets.0.conv_shortcut MXFP4 W4A4 640ร—320ร—1ร—1 320 178,432 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—256], W_scales U8 [640ร—16], bias BF16 [640], pre_scale BF16 [512]
down_blocks.2.attentions.0.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.0.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.1.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.1.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.resnets.0.conv_shortcut MXFP4 W4A4 1280ร—640ร—1ร—1 640 528,384 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], bias BF16 [1280], pre_scale BF16 [768]
mid_block.attentions.0.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
mid_block.attentions.0.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.0.resnets.0.conv_shortcut MXFP4 W4A4 1280ร—2560ร—1ร—1 2,560 1,750,528 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1280], W_scales U8 [1280ร—80], bias BF16 [1280], pre_scale BF16 [2560]
up_blocks.0.resnets.1.conv_shortcut MXFP4 W4A4 1280ร—2560ร—1ร—1 2,560 1,750,528 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1280], W_scales U8 [1280ร—80], bias BF16 [1280], pre_scale BF16 [2560]
up_blocks.0.resnets.2.conv_shortcut MXFP4 W4A4 1280ร—2560ร—1ร—1 2,560 1,750,528 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1280], W_scales U8 [1280ร—80], bias BF16 [1280], pre_scale BF16 [2560]
up_blocks.1.attentions.0.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.0.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.1.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.1.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.2.proj_in MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.2.proj_out MXFP4 W4A4 1280ร—1280ร—1ร—1 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.resnets.0.conv_shortcut MXFP4 W4A4 1280ร—2560ร—1ร—1 2,560 1,750,528 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1280], W_scales U8 [1280ร—80], bias BF16 [1280], pre_scale BF16 [2560]
up_blocks.1.resnets.1.conv_shortcut MXFP4 W4A4 1280ร—2560ร—1ร—1 2,560 1,750,528 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1280], W_scales U8 [1280ร—80], bias BF16 [1280], pre_scale BF16 [2560]
up_blocks.1.resnets.2.conv_shortcut MXFP4 W4A4 1280ร—1920ร—1ร—1 1,920 1,401,344 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—1024], W_scales U8 [1280ร—64], bias BF16 [1280], pre_scale BF16 [2048]
up_blocks.2.attentions.0.proj_in MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.0.proj_out MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.1.proj_in MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.1.proj_out MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.2.proj_in MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.2.proj_out MXFP4 W4A4 640ร—640ร—1ร—1 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.resnets.0.conv_shortcut MXFP4 W4A4 640ร—1920ร—1ร—1 1,920 703,744 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1024], W_scales U8 [640ร—64], bias BF16 [640], pre_scale BF16 [2048]
up_blocks.2.resnets.1.conv_shortcut MXFP4 W4A4 640ร—1280ร—1ร—1 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
up_blocks.2.resnets.2.conv_shortcut MXFP4 W4A4 640ร—960ร—1ร—1 960 353,536 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—512], W_scales U8 [640ร—32], bias BF16 [640], pre_scale BF16 [1024]
up_blocks.3.attentions.0.proj_in MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.0.proj_out MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.1.proj_in MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.1.proj_out MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.2.proj_in MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.2.proj_out MXFP4 W4A4 320ร—320ร—1ร—1 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.resnets.0.conv_shortcut MXFP4 W4A4 320ร—960ร—1ร—1 960 178,816 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—512], W_scales U8 [320ร—32], bias BF16 [320], pre_scale BF16 [1024]
up_blocks.3.resnets.1.conv_shortcut MXFP4 W4A4 320ร—640ร—1ร—1 640 134,784 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], bias BF16 [320], pre_scale BF16 [768]
up_blocks.3.resnets.2.conv_shortcut MXFP4 W4A4 320ร—640ร—1ร—1 640 134,784 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], bias BF16 [320], pre_scale BF16 [768]
Attention projections ยท 128 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
mid_block.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
mid_block.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_k MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_v MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_k MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_q MXFP4 W4A4 1280ร—1280 1,280 875,008 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_v MXFP4 W4A4 1280ร—768 768 525,824 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—384], W_scales U8 [1280ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_k MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_v MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_k MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 640ร—640 640 265,984 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], bias BF16 [640], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_q MXFP4 W4A4 640ร—640 640 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_v MXFP4 W4A4 640ร—768 768 264,704 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—384], W_scales U8 [640ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_k MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_v MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_k MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_v MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_k MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_v MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_k MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_v MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_k MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_v MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_k MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_out.0 MXFP4 W4A4 320ร—320 320 90,752 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], bias BF16 [320], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_q MXFP4 W4A4 320ร—320 320 90,112 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—256], W_scales U8 [320ร—16], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_v MXFP4 W4A4 320ร—768 768 134,144 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—384], W_scales U8 [320ร—24], pre_scale BF16 [768]
MLP projections ยท 32 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
down_blocks.0.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 2560ร—320 320 704,512 B_inv_T BF16 [32ร—32], W_packed U8 [2560ร—256], W_scales U8 [2560ร—16], bias BF16 [2560], pre_scale BF16 [512]
down_blocks.0.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
down_blocks.0.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 2560ร—320 320 704,512 B_inv_T BF16 [32ร—32], W_packed U8 [2560ร—256], W_scales U8 [2560ร—16], bias BF16 [2560], pre_scale BF16 [512]
down_blocks.0.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
down_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 5120ร—640 640 2,102,784 B_inv_T BF16 [32ร—32], W_packed U8 [5120ร—384], W_scales U8 [5120ร—24], bias BF16 [5120], pre_scale BF16 [768]
down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 640ร—2560 2,560 878,848 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1280], W_scales U8 [640ร—80], bias BF16 [640], pre_scale BF16 [2560]
down_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 5120ร—640 640 2,102,784 B_inv_T BF16 [32ร—32], W_packed U8 [5120ร—384], W_scales U8 [5120ร—24], bias BF16 [5120], pre_scale BF16 [768]
down_blocks.1.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 640ร—2560 2,560 878,848 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1280], W_scales U8 [640ร—80], bias BF16 [640], pre_scale BF16 [2560]
down_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
down_blocks.2.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
down_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
down_blocks.2.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
mid_block.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
mid_block.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
up_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
up_blocks.1.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
up_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
up_blocks.1.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
up_blocks.1.attentions.2.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 10240ร—1280 1,280 6,988,288 B_inv_T BF16 [32ร—32], W_packed U8 [10240ร—640], W_scales U8 [10240ร—40], bias BF16 [10240], pre_scale BF16 [1280]
up_blocks.1.attentions.2.transformer_blocks.0.ff.net.2 MXFP4 W4A4 1280ร—5120 5,120 3,496,448 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—2560], W_scales U8 [1280ร—160], bias BF16 [1280], pre_scale BF16 [5120]
up_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 5120ร—640 640 2,102,784 B_inv_T BF16 [32ร—32], W_packed U8 [5120ร—384], W_scales U8 [5120ร—24], bias BF16 [5120], pre_scale BF16 [768]
up_blocks.2.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 640ร—2560 2,560 878,848 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1280], W_scales U8 [640ร—80], bias BF16 [640], pre_scale BF16 [2560]
up_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 5120ร—640 640 2,102,784 B_inv_T BF16 [32ร—32], W_packed U8 [5120ร—384], W_scales U8 [5120ร—24], bias BF16 [5120], pre_scale BF16 [768]
up_blocks.2.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 640ร—2560 2,560 878,848 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1280], W_scales U8 [640ร—80], bias BF16 [640], pre_scale BF16 [2560]
up_blocks.2.attentions.2.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 5120ร—640 640 2,102,784 B_inv_T BF16 [32ร—32], W_packed U8 [5120ร—384], W_scales U8 [5120ร—24], bias BF16 [5120], pre_scale BF16 [768]
up_blocks.2.attentions.2.transformer_blocks.0.ff.net.2 MXFP4 W4A4 640ร—2560 2,560 878,848 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—1280], W_scales U8 [640ร—80], bias BF16 [640], pre_scale BF16 [2560]
up_blocks.3.attentions.0.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 2560ร—320 320 704,512 B_inv_T BF16 [32ร—32], W_packed U8 [2560ร—256], W_scales U8 [2560ร—16], bias BF16 [2560], pre_scale BF16 [512]
up_blocks.3.attentions.0.transformer_blocks.0.ff.net.2 MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
up_blocks.3.attentions.1.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 2560ร—320 320 704,512 B_inv_T BF16 [32ร—32], W_packed U8 [2560ร—256], W_scales U8 [2560ร—16], bias BF16 [2560], pre_scale BF16 [512]
up_blocks.3.attentions.1.transformer_blocks.0.ff.net.2 MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
up_blocks.3.attentions.2.transformer_blocks.0.ff.net.0.proj MXFP4 W4A4 2560ร—320 320 704,512 B_inv_T BF16 [32ร—32], W_packed U8 [2560ร—256], W_scales U8 [2560ร—16], bias BF16 [2560], pre_scale BF16 [512]
up_blocks.3.attentions.2.transformer_blocks.0.ff.net.2 MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
Timestep and other linears ยท 24 layers
Layer Precision Original weight shape Hessian width Stored bytes Tensor dtypes and shapes
down_blocks.0.resnets.0.time_emb_proj MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
down_blocks.0.resnets.1.time_emb_proj MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
down_blocks.1.resnets.0.time_emb_proj MXFP4 W4A4 640ร—1280 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
down_blocks.1.resnets.1.time_emb_proj MXFP4 W4A4 640ร—1280 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
down_blocks.2.resnets.0.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.2.resnets.1.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.3.resnets.0.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
down_blocks.3.resnets.1.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
mid_block.resnets.0.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
mid_block.resnets.1.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
time_embedding.linear_1 MXFP4 W4A4 1280ร—320 320 353,792 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—256], W_scales U8 [1280ร—16], bias BF16 [1280], pre_scale BF16 [512]
time_embedding.linear_2 MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.0.resnets.0.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.0.resnets.1.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.0.resnets.2.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.resnets.0.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.resnets.1.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.1.resnets.2.time_emb_proj MXFP4 W4A4 1280ร—1280 1,280 877,568 B_inv_T BF16 [32ร—32], W_packed U8 [1280ร—640], W_scales U8 [1280ร—40], bias BF16 [1280], pre_scale BF16 [1280]
up_blocks.2.resnets.0.time_emb_proj MXFP4 W4A4 640ร—1280 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
up_blocks.2.resnets.1.time_emb_proj MXFP4 W4A4 640ร—1280 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
up_blocks.2.resnets.2.time_emb_proj MXFP4 W4A4 640ร—1280 1,280 441,088 B_inv_T BF16 [32ร—32], W_packed U8 [640ร—640], W_scales U8 [640ร—40], bias BF16 [640], pre_scale BF16 [1280]
up_blocks.3.resnets.0.time_emb_proj MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
up_blocks.3.resnets.1.time_emb_proj MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]
up_blocks.3.resnets.2.time_emb_proj MXFP4 W4A4 320ร—1280 1,280 222,848 B_inv_T BF16 [32ร—32], W_packed U8 [320ร—640], W_scales U8 [320ร—40], bias BF16 [320], pre_scale BF16 [1280]

Run locally

Use CUDA PyTorch with Triton, Diffusers 0.39.0, Transformers 4.57.6, Accelerate, Safetensors, and Hugging Face Hub. The packed U-Net uses the included custom loader. Use backend="native" for native MXFP4 on a compatible Blackwell setup, or backend="emulated" to reproduce the L4 evaluation. Native inference was outside the reported test.

import sys
from pathlib import Path

import torch
from huggingface_hub import snapshot_download

directory = Path(snapshot_download("ApacheOne/SD15_MACaQ_pruned_fullwidth"))
sys.path.insert(0, str(directory))
from load_pipeline import load_pipeline

torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
pipe = load_pipeline(directory, device="cuda", backend="emulated")
image = pipe(
    "A ceramic teapot on a wooden table",
    height=512, width=512,
    num_inference_steps=4, guidance_scale=1.2,
    generator=torch.Generator("cuda").manual_seed(7003),
).images[0]
image.save("macaq_fullwidth.png")

Sources and terms

Source checkpoint SHA-256 is 1a16a9524427019dc513b398087035e7e36ad2f2aa9691bd71d28ad959e94180.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support