Skip to content

v4.1: Qualcomm models

Choose a tag to compare

View all tags
@anhappdev anhappdev released this 10 Oct 06:28
v4.1-qualcomm
1f99633
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Verified
Learn about vigilant mode. 
MD5 (betas.bin) = 09d2e4306d319caf1b34e6afb5c63c22
MD5 (lambdas.bin) = c7179725ec31a6e2c7daf008a5e1ff23
MD5 (sd_precompute_data.tar) = beb7fe2da40042fb585bb8cb95d86b4d
MD5 (text_encoder.serialized.bin) = 6da7b95fa467e99af2b9f80c7afe3734
MD5 (unet.serialized.bin) = 3b504b92cbd788d713ca9cfc5b19d596
MD5 (vae_decoder.serialized.bin) = c7762e64c2596abe7f16614709cc5482

MD5 (mobile_mosaic_htp.dlc) = 3c0dfbacda053773d6afb34503d9991a
MD5 (mobilebert_quantized_htp.dlc) = 96d947175f04950898a372890907dda1
MD5 (mobilebert_quantized_htp_O2.dlc) = f8631dbd69819438d6b317c204fa80d7
MD5 (mobilenet_v4_htp.dlc) = 56e5039260e20e5c2a0b54cc0fac8098
MD5 (mobilenet_v4_htp_batched_4.dlc) = 7863deea588936fe6e09565ed47dde95
MD5 (mobilenet_v4_htp_batched_4_O2.dlc) = 80ba82f2a628ab712d812d06524d2bd8
MD5 (snusr_htp.dlc) = 668da9816073d67972704e237137a50f
MD5 (snusr_htp_O2.dlc) = 76b33f02ebfa6294a0e973aaf91116fa
MD5 (ssd_mobiledet_qat_htp.dlc) = 49c6afbfefffb78269fe73a6ee1b4a85

Note from Qualcomm for Stable Diffusion v1.5:
AI Model Efficiency Toolkit (AIMET) is used to create quantization simulation models (QuantSim) for the text encoder, U-Net, and VAE using a mixed-precision quantization scheme. A calibration process is used to create these QuantSim models where per-layer quantization encodings are determined using representative data samples. The QuantSim models simulate running the stable diffusion models on a quantized target. In addition, for the Text Encoder model, the AIMET Adaptive Rounding (AdaRound) technique is applied to get a boost in quantized accuracy.
Time Embeddings are precomputed during quantization step above. During inference, the model for Unet takes 1x1280 precomputed Time Embeddings.

Assets 18
Loading