dynamic_vram: fix pinning with model defined dtype#12208
Merged
comfyanonymous merged 1 commit intoComfy-Org:masterfrom Feb 1, 2026
Merged
Conversation
pinned memory was converted back to pinning the CPU side weight without any changes. Fix the pinner to use the CPU weight and not the model defined geometry. This will either save RAM or stop buffer overruns when the types mismatch. Fix the model defined weight caster to use the [ s.weight, s.bias ] interpretation, as xfer_dest might be the flattened pin now. Fix the detection of needing to cast to not be conditional on !pin.
rattus128
requested review from
Kosinkadink,
comfyanonymous and
guill
as code owners
|
This still did not fix the issue with the division by 4 error for me. Reverting Comfy to 873de5f makes it work again |
Contributor
Author
It's a different issue to the bug report, but thanks for the log! What is the safetensors file it is loading? Your trace suggests a misaligned sft weight which we can fix, but helps if I have a handle on the safetensors file. |
|
it is an SDXL checkpoint, here are the safetensors meta: {
"format": "pt",
"sd_merge_recipe": {
"type": "webui",
"primary_model_hash": "d9d414bd112afb0db7363bb7b976da34d840d80e7d1ba3b3943c0dfbeb70499c",
"secondary_model_hash": null,
"tertiary_model_hash": null,
"interp_method": "Weighted sum",
"multiplier": 0.3,
"save_as_half": false,
"custom_name": "ilustmix_v60",
"config_source": 0,
"bake_in_vae": "sdxlVAE_sdxlVAE_264202.safetensors",
"discard_weights": "",
"is_inpainting": false,
"is_instruct_pix2pix": false
},
"sd_merge_models": {
"d9d414bd112afb0db7363bb7b976da34d840d80e7d1ba3b3943c0dfbeb70499c": {
"name": "ilustmix_v40.safetensors",
"legacy_hash": "a56eb33a",
"sd_merge_recipe": {
"type": "webui",
"primary_model_hash": "cb3a162f3af00b00dc500e815953d5983dc33a4be3c480a855a27ff4db31caac",
"secondary_model_hash": "a3a562d381e7c54f326e5f68a280a660b0f76571033ec625c90b64a9ea29066f",
"tertiary_model_hash": null,
"interp_method": "Weighted sum",
"multiplier": 0.4,
"save_as_half": false,
"custom_name": "ilustmix_v40",
"config_source": 0,
"bake_in_vae": "sdxlVAE_sdxlVAE_264202.safetensors",
"discard_weights": "",
"is_inpainting": false,
"is_instruct_pix2pix": false
}
},
"cb3a162f3af00b00dc500e815953d5983dc33a4be3c480a855a27ff4db31caac": {
"name": "ilustmix_v30_1357311.safetensors",
"legacy_hash": "88426183",
"sd_merge_recipe": {
"type": "webui",
"primary_model_hash": "1620b6d4794f996d38dae6b0b522d05a2e92d5c938791d86973bbf404c9b0b7d",
"secondary_model_hash": null,
"tertiary_model_hash": null,
"interp_method": "Weighted sum",
"multiplier": 0.2,
"save_as_half": false,
"custom_name": "ilustmix",
"config_source": 0,
"bake_in_vae": "None",
"discard_weights": "",
"is_inpainting": false,
"is_instruct_pix2pix": false
}
},
"1620b6d4794f996d38dae6b0b522d05a2e92d5c938791d86973bbf404c9b0b7d": {
"name": "ilustmix_v20_1221746.safetensors",
"legacy_hash": "97fb5c13",
"sd_merge_recipe": {
"type": "webui",
"primary_model_hash": "cbd74f1aeb3a6f9c6d0bca71c658a9fb01f6e192a9027f5d099b986ea0950f64",
"secondary_model_hash": "7a48fb8eae2bbb1b4ecdfe479fdede83bf252f4841cc1f2dde278c6c39636b91",
"tertiary_model_hash": null,
"interp_method": "Weighted sum",
"multiplier": 0.5,
"save_as_half": false,
"custom_name": "ilustmix_v20",
"config_source": 0,
"bake_in_vae": "None",
"discard_weights": "",
"is_inpainting": false,
"is_instruct_pix2pix": false
}
},
"cbd74f1aeb3a6f9c6d0bca71c658a9fb01f6e192a9027f5d099b986ea0950f64": {
"name": "illustrij_v50_1153542.safetensors",
"legacy_hash": "4cb125f4",
"sd_merge_recipe": null
},
"7a48fb8eae2bbb1b4ecdfe479fdede83bf252f4841cc1f2dde278c6c39636b91": {
"name": "cyberillustrious_v10_1169622.safetensors",
"legacy_hash": "683959ce",
"sd_merge_recipe": null
},
"3a2df45ef8a243818be8aaa312dee2e187bfe6851a9cc6b57022b3a1440279bb": {
"name": "prefectIllustriousXL_v10_1282816.safetensors",
"legacy_hash": "832a96e1",
"sd_merge_recipe": null
},
"a3a562d381e7c54f326e5f68a280a660b0f76571033ec625c90b64a9ea29066f": {
"name": "illustrij_v10_1441661.safetensors",
"legacy_hash": "c0656ffa",
"sd_merge_recipe": null
},
"e10c2b912e3adeff666e8e56d44b1e9d294cb04bff335b552a63d00f58a0346b": {
"name": "jedpointreal_v1ILVae_1541657.safetensors",
"legacy_hash": "288f91ff",
"sd_merge_recipe": null
}
}
}just in case my start up options, but I do not think they cause the issue: python venv |
simonri
pushed a commit
to simonri/ComfyUI-flash-attention-3
that referenced
this pull request
Feb 2, 2026
pinned memory was converted back to pinning the CPU side weight without any changes. Fix the pinner to use the CPU weight and not the model defined geometry. This will either save RAM or stop buffer overruns when the types mismatch. Fix the model defined weight caster to use the [ s.weight, s.bias ] interpretation, as xfer_dest might be the flattened pin now. Fix the detection of needing to cast to not be conditional on !pin.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
pinned memory was converted back to pinning the CPU side weight without any changes. Fix the pinner to use the CPU weight and not the model defined geometry. This will either save RAM or stop buffer overruns when the types mismatch.
Fix the model defined weight caster to use the [ s.weight, s.bias ] interpretation, as xfer_dest might be the flattened pin now. Fix the detection of needing to cast to not be conditional on !pin.
#12205
Example Test conditions:
RTX3060, Linux, --fast dynamic_vram
LTX2 VAE decode with on-demand model:
Before:
First crash:
Second crash (matches issue OP):
After: