You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
MiniMax H3: aimdo memory compile error 'could not start recording' on a single RTX 5090 with a stock text-to-video graph (v0.35.2, comfy-aimdo 0.5.3) #16342
On a single GPU, a plain MiniMax H3 text-to-video graph (no ControlNet, no
sparse attention, no LoRA) fails at SamplerCustomAdvanced with:
RuntimeError: aimdo memory compile error: could not start recording
--disable-comfy-compiler makes the same graph run. This looks distinct from #16223, which is a hostbuf_read_file_slice device-copy failure when two
instances stage models on two GPUs concurrently; here there is one instance,
one GPU, and the recorder itself does not start.
Environment
ComfyUI v0.35.2, Windows 11
comfy-aimdo 0.5.3, comfy-kitchen 0.2.33 (the versions pinned by 0.35.2's requirements.txt)
torch 2.9.1+cu130, Python 3.13
RTX 5090, single instance, launched as python main.py --port 8188 (no --enable-dynamic-vram, no pinned-memory flags)
Models: minimax_h3_fl2va_pruned_int8_convrot, qwen3vl_32b_minimax_h3_nvfp4_awq, H3 video and audio VAEs
Queue. The sampler node raises the error above before the first step completes.
Relaunch with --disable-comfy-compiler; the same graph completes.
The same error appears with the core MiniMaxH3FunControlNetApply in the graph and with a third-party ControlNet in the graph, so it is not specific to model patches.
Notes
I do not have a second machine to say whether this is cu130-specific or 5090-specific. Happy to run anything that helps narrow it: verbose aimdo logging, a different driver, or a smaller resolution.
This one is intermittent: the identical graph ran to completion twice and failed once, at the sampler, while streaming weights. It happens with --disable-comfy-compiler set, so it is the host-buffer streaming path rather than the recorder. Same environment as above (v0.35.2, comfy-aimdo 0.5.3, RTX 5090, Windows 11, torch 2.9.1+cu130). The graph in question was MiniMax H3 with a model patch applied, but the earlier "could not start recording" failure reproduces on a stock graph, so I would not read the patch as the cause here either.
The second trace is useful because it separates the recorder failure from the host-buffer streaming failure.
I’d run a minimal 2×2 matrix from fresh processes on the stock graph:
default
--disable-comfy-compiler
--disable-dynamic-vram
both disabled
Before those runs, capture rather than modify any inherited CUDA_VISIBLE_DEVICES, GPU_DEVICE_ORDINAL, NVIDIA_VISIBLE_DEVICES, etc. #16223 ultimately turned out to have stale device-pinning variables in the process environment, so it’s worth ruling that out here before treating the similar host-buffer trace as the same aimdo defect.
If compiler-off eliminates only could not start recording, while DynamicVRAM-off eliminates only HostBuffer.read_file_slice, that would strongly suggest two separate regressions.
If the host-buffer failure still occurs with DynamicVRAM disabled, I’d start looking lower at the allocator/driver path rather than the recorder.
Getting the same issue today although it only began mid way through the day after doing many gens and it actually happened with LTX 2.5. I tried Minimax H3 after and the same thing happened...
Updated again just now, hopefully it is resolved but just notifying the same issue here.
The fact that this appeared after many successful generations and then reproduced across both LTX 2.5 and H3 is useful.
That makes an H3-specific workflow problem less likely and raises the possibility of process-lifetime state accumulating somewhere in the staging / allocator / host-buffer path.
If it happens again after the latest update, I’d record:
whether a full ComfyUI process restart immediately clears it
roughly how many successful generations or how much elapsed time passes before the first failure
which models were loaded/offloaded before it starts
VRAM and system committed memory just before the first bad run
Then rerun the earlier compiler/DynamicVRAM 2×2 from a fresh process, rather than after the session is already contaminated.
If a restart reliably clears an LTX → H3 sequence but changing workflows/models does not, I’d investigate persistent staged-model / host-buffer / allocator state before looking at either model graph.
After a few successful H3 generations in 1MP, I increased it to 2MP and it started failing with RuntimeError: aimdo memory compile error: could not start recording. I thought it was related to the resolution, so I brought it down back to 1MP but the error kept happening.
Unfortunately I no longer have the workflow, but it was basically the standard H3 ref2vid with Model Attention Backend and KJ's Model Preview Override nodes. I bypassed both but to no avail.
After restarting ComfyUI, I was able to generate again, including in 2MP.
Cross-reference from a related but different failure on the same component (not a duplicate: this one is malloc_graph_pause() failing mid-sampling, not the recorder failing to start).
Setup: ComfyUI 0.42.3, comfy-aimdo 0.5.6, RTX 4090 (48 GB), MiniMax-H3 + Spark-H3 block-sparse attention + a two-pass sampler with 4-tile high-res tiling. The prompt dies deterministically (7/7 runs) at the same DiT block with RuntimeError: aimdo memory compile error, raised from malloc_graph_pause().
Cause (traced in comfy-aimdo source, cross-checked against the 0.5.6 binary): malloc_graph_pop() severs every allocation that is still alive in the popped scope and registers it as a rogue candidate; finalize_rogues() only adopts candidates at the root pop. If the application frees one of those pointers before that, graph_failed() sets g->failed, and from then on every alloc/free/pause/push/pop/set_stream returns false — the graph is dead for the rest of the process and there is no reset path.
Filed with source references, runtime evidence (a probe reading the native failed byte at handle+0x5E0D0, which flips from 0 to 1 between two block-boundary calls) and a 27-line patch: Comfy-Org/comfy-aimdo#129 (PR Comfy-Org/comfy-aimdo#130).
I am not claiming this is the same bug as the could not start recording variant. But if a repro survives --disable-comfy-compiler and then fails later in the same session, this is worth ruling out. For us --disable-comfy-compiler was the only workaround that held before the patch.
Activity
wyzborrero commentedon Sep 15, 2026
A second aimdo symptom from the same machine, single GPU, which may be the same root cause as #16223 without the two-instance setup:
This one is intermittent: the identical graph ran to completion twice and failed once, at the sampler, while streaming weights. It happens with
--disable-comfy-compilerset, so it is the host-buffer streaming path rather than the recorder. Same environment as above (v0.35.2, comfy-aimdo 0.5.3, RTX 5090, Windows 11, torch 2.9.1+cu130). The graph in question was MiniMax H3 with a model patch applied, but the earlier "could not start recording" failure reproduces on a stock graph, so I would not read the patch as the cause here either.InterAuc commentedon Sep 16, 2026
The second trace is useful because it separates the recorder failure from the host-buffer streaming failure.
I’d run a minimal 2×2 matrix from fresh processes on the stock graph:
--disable-comfy-compiler--disable-dynamic-vramBefore those runs, capture rather than modify any inherited
CUDA_VISIBLE_DEVICES,GPU_DEVICE_ORDINAL,NVIDIA_VISIBLE_DEVICES, etc. #16223 ultimately turned out to have stale device-pinning variables in the process environment, so it’s worth ruling that out here before treating the similar host-buffer trace as the same aimdo defect.If compiler-off eliminates only
could not start recording, while DynamicVRAM-off eliminates onlyHostBuffer.read_file_slice, that would strongly suggest two separate regressions.If the host-buffer failure still occurs with DynamicVRAM disabled, I’d start looking lower at the allocator/driver path rather than the recorder.
hicaonmb commentedon Sep 18, 2026
how to resolve this problem? I met the same error on 2080ti 22g when i updated comfyui to latest
lebakerino commentedon Sep 18, 2026
Getting the same issue today although it only began mid way through the day after doing many gens and it actually happened with LTX 2.5. I tried Minimax H3 after and the same thing happened...
Updated again just now, hopefully it is resolved but just notifying the same issue here.
InterAuc commentedon Sep 19, 2026
The fact that this appeared after many successful generations and then reproduced across both LTX 2.5 and H3 is useful.
That makes an H3-specific workflow problem less likely and raises the possibility of process-lifetime state accumulating somewhere in the staging / allocator / host-buffer path.
If it happens again after the latest update, I’d record:
Then rerun the earlier compiler/DynamicVRAM 2×2 from a fresh process, rather than after the session is already contaminated.
If a restart reliably clears an LTX → H3 sequence but changing workflows/models does not, I’d investigate persistent staged-model / host-buffer / allocator state before looking at either model graph.
fappaz commentedon Sep 29, 2026
Definitely not 5090-specific, I got the same error on my 5060Ti:
After a few successful H3 generations in 1MP, I increased it to 2MP and it started failing with
RuntimeError: aimdo memory compile error: could not start recording. I thought it was related to the resolution, so I brought it down back to 1MP but the error kept happening.Unfortunately I no longer have the workflow, but it was basically the standard H3 ref2vid with
Model Attention Backendand KJ'sModel Preview Overridenodes. I bypassed both but to no avail.After restarting ComfyUI, I was able to generate again, including in 2MP.
lonelystarCX commentedon Oct 22, 2026
Cross-reference from a related but different failure on the same component (not a duplicate: this one is
malloc_graph_pause()failing mid-sampling, not the recorder failing to start).Setup: ComfyUI 0.42.3, comfy-aimdo 0.5.6, RTX 4090 (48 GB), MiniMax-H3 + Spark-H3 block-sparse attention + a two-pass sampler with 4-tile high-res tiling. The prompt dies deterministically (7/7 runs) at the same DiT block with
RuntimeError: aimdo memory compile error, raised frommalloc_graph_pause().Cause (traced in comfy-aimdo source, cross-checked against the 0.5.6 binary):
malloc_graph_pop()severs every allocation that is still alive in the popped scope and registers it as a rogue candidate;finalize_rogues()only adopts candidates at the root pop. If the application frees one of those pointers before that,graph_failed()setsg->failed, and from then on everyalloc/free/pause/push/pop/set_streamreturns false — the graph is dead for the rest of the process and there is no reset path.Filed with source references, runtime evidence (a probe reading the native
failedbyte athandle+0x5E0D0, which flips from 0 to 1 between two block-boundary calls) and a 27-line patch: Comfy-Org/comfy-aimdo#129 (PR Comfy-Org/comfy-aimdo#130).I am not claiming this is the same bug as the
could not start recordingvariant. But if a repro survives--disable-comfy-compilerand then fails later in the same session, this is worth ruling out. For us--disable-comfy-compilerwas the only workaround that held before the patch.fappaz commentedon Nov 2, 2026
Just a note that I updated ComfyUI recently but I still sporadically get the same error:
My specs:
CaLvin713 commentedon Nov 15, 2026
never got this error before... updated and now getting it :
[ERROR] !!! Exception during processing !!! aimdo memory compile error: could not start recording
[ERROR] Traceback (most recent call last):
File "C:\comfy60\ComfyUI\execution.py", line 556, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\execution.py", line 353, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\execution.py", line 327, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "C:\comfy60\ComfyUI\execution.py", line 315, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "C:\comfy60\ComfyUI\nodes.py", line 1633, in sample
return common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\nodes.py", line 1597, in common_ksampler
samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\sample.py", line 82, in sample
samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1464, in sample
return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1354, in sample
return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1336, in sample
output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1263, in outer_sample
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1238, in inner_sample
samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1005, in sample
samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\python_embeded\Lib\site-packages\torch\utils_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\k_diffusion\sampling.py", line 205, in sample_euler
denoised = model(x, sigma_hat * s_in, **extra_args)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 640, in call
out = self.inner_model(x, sigma, model_options=model_options, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1208, in call
return self.outer_predict_noise(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1215, in outer_predict_noise
).execute(x, timestep, model_options, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 1218, in predict_noise
return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 620, in sampling_function
out = calc_cond_batch(model, conds, x, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 211, in calc_cond_batch
return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 219, in _calc_cond_batch_outer
return executor.execute(model, conds, x_in, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\samplers.py", line 335, in calc_cond_batch
output = model.apply_model(input_x, timestep, **c).chunk(batch_chunks)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\model_base.py", line 215, in apply_model
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\model_base.py", line 259, in _apply_model
model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1778, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1789, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\ldm\qwen_image21\model.py", line 420, in forward
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\ComfyUI\comfy\ldm\qwen_image21\model.py", line 470, in _forward
comfy.model_prefetch.malloc_graph_begin(x.device)
File "C:\comfy60\ComfyUI\comfy\model_prefetch.py", line 77, in malloc_graph_begin
graph = comfy_aimdo.malloc_graph.record(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfy60\python_embeded\Lib\site-packages\comfy_aimdo\malloc_graph.py", line 84, in record
raise RuntimeError("aimdo memory compile error: could not start recording")
RuntimeError: aimdo memory compile error: could not start recording
System Info
Arguments: ComfyUI\main.py --windows-standalone-build --vram-headroom 2 --max-upload-size 2024
Embedded Python: true
Templates Version: 0.11.107
OS: win32
Python Version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
PyTorch Version: 2.13.0+cu130
RAM Free: 42.67 GB
RAM Total: 61.65 GB
Devices
VRAM Total: 31.82 GB
VRAM Free: 28.15 GB
Torch VRAM Total: 416 MB
Torch VRAM Free: 138.16 MB