Performance on Mac Mini M4 / 24Gb Ram

[timbo1975]

Hi Ace-Step Team, firstly thank you for your awesome work this project is really amazing! I've seen a few YT videos running on Windows machines with 4090Ti cards etc and was wondering if anyone has benchmarked on a Mac Mini M4? I have it running and I am willing to test, at the moment it takes around 5 mins to generate a song with lyrics does that seem right?! I have it running with the following parameters;

--bf16 false --overlapped_decode true

[timbo1975]

For anyone who might find this useful;

I have it running pretty well acestep --torch_compile true --overlapped_decode true --bf16 false. I have also tried --bf16 true and didn't notice much difference. Unfortunately using torch_compile it crashes on second generation and seems to be related to pytorch rather than ACE-Step, that said it's quicker to restart the server than use it without torch_compile (slower). My ram usage is way higher than stated on the examples up at 21gb and using swap.

My parameters are as follows;

Defaults for everything but Inferences Steps are 27 and duration is 45 seconds and generations take about 1-2 mins to appear in the UI.

[ChuxiJ]

can you help me to create an issue to apple mlx team, it may have better suport for this model

[timbo1975]

I'm happy to help I'm not familiar with raising issues (let me know where) but have just installed the latest release ACE-Step from you and the team and the crash seems to be resolved. I have python running in a venv so I know it's clean - I only installed ACE-Step (love it!). Let me know the best place to post any things I find that will help you with the project.

UPDATE: It happened again here is the log.

C0513 23:37:17.109000 18675 torch/_inductor/scheduler.py:1204] [0/1] Error in codegen for ComputedBuffer(name='buf5', layout=FixedLayout('mps:0', torch.float32, size=[1, 12, s0, s0], stride=[12*s02, s02, s0, 1]), data=Pointwise(device=device(type='mps', index=0), dtype=torch.float32, inner_fn=<function make_pointwise..inner..inner_fn at 0x311e39940>, ranges=[1, 12, s0, s0]))
Traceback (most recent call last):
File ".../venv/lib/python3.11/site-packages/gradio/queueing.py", line 625, in process_events
response = await route_utils.call_process_api(
File ".../venv/lib/python3.11/site-packages/gradio/route_utils.py", line 322, in call_process_api
output = await app.get_blocks().process_api(
File ".../venv/lib/python3.11/site-packages/gradio/blocks.py", line 2137, in process_api
result = await self.call_function(
File ".../venv/lib/python3.11/site-packages/gradio/blocks.py", line 1663, in call_function
prediction = await anyio.to_thread.run_sync( # type: ignore
File ".../venv/lib/python3.11/site-packages/anyio/to_thread.py", line 56, in run_sync
return await get_async_backend().run_sync_in_worker_thread(
File ".../venv/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 2470, in run_sync_in_worker_thread
return await future
File ".../venv/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 967, in run
result = context.run(func, *args)
File ".../venv/lib/python3.11/site-packages/gradio/utils.py", line 890, in wrapper
response = f(*args, **kwargs)
File ".../acestep/ui/components.py", line 773, in extend_process_func
return text2music_process_func(
File ".../acestep/pipeline_ace_step.py", line 1649, in call
encoder_text_hidden_states, text_attention_mask = self.get_text_embeddings(
File ".../acestep/cpu_offload.py", line 32, in wrapper
return func(self, *args, **kwargs)
File ".../acestep/pipeline_ace_step.py", line 531, in get_text_embeddings
outputs = self.text_encoder_model(**inputs)
File ".../venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File ".../venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
File ".../venv/lib/python3.11/site-packages/torch/_dynamo/eval_frame.py", line 663, in _fn
raise e.remove_dynamo_frames() from None
File ".../venv/lib/python3.11/site-packages/torch/_inductor/compile_fx.py", line 760, in _compile_fx_inner
raise InductorError(e, currentframe()).with_traceback(
File ".../venv/lib/python3.11/site-packages/torch/_inductor/compile_fx.py", line 745, in _compile_fx_inner
mb_compiled_graph = fx_codegen_and_compile(
File ".../venv/lib/python3.11/site-packages/torch/_inductor/compile_fx.py", line 1295, in fx_codegen_and_compile
return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/compile_fx.py", line 1197, in codegen_and_compile
compiled_fn = graph.compile_to_module().call
File ".../venv/lib/python3.11/site-packages/torch/_inductor/graph.py", line 2083, in compile_to_module
return self._compile_to_module()
File ".../venv/lib/python3.11/site-packages/torch/_inductor/graph.py", line 2091, in _compile_to_module
self.codegen_with_cpp_wrapper() if self.cpp_wrapper else self.codegen()
File ".../venv/lib/python3.11/site-packages/torch/_inductor/graph.py", line 2002, in codegen
self.scheduler.codegen()
File ".../venv/lib/python3.11/site-packages/torch/_inductor/scheduler.py", line 4135, in codegen
else self._codegen(self.nodes)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/scheduler.py", line 4264, in _codegen
self.get_backend(device).codegen_node(node)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 1320, in codegen_node
return self.codegen_node_schedule(
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 1361, in codegen_node_schedule
self.codegen_node_schedule_with_kernel(node_schedule, kernel)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 1461, in codegen_node_schedule_with_kernel
node.codegen(index_vars)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/scheduler.py", line 1202, in codegen
self._body(*index_vars)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/loop_body.py", line 407, in call
result = self.root_block()
File ".../venv/lib/python3.11/site-packages/torch/_inductor/loop_body.py", line 476, in call
return InterpreterShim(graph, submodules).run(V.get_ops_handler())
File ".../venv/lib/python3.11/site-packages/torch/_inductor/loop_body.py", line 60, in run
return super().run(*args, **kwargs)
File ".../venv/lib/python3.11/site-packages/torch/fx/interpreter.py", line 171, in run
self.env[node] = self.run_node(node)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/loop_body.py", line 56, in run_node
return super().run_node(n)
File ".../venv/lib/python3.11/site-packages/torch/fx/interpreter.py", line 240, in run_node
return getattr(self, n.op)(n.target, args, kwargs)
File ".../venv/lib/python3.11/site-packages/torch/fx/interpreter.py", line 344, in call_method
return getattr(self_obj, target)(*args_tail, **kwargs)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/sizevars.py", line 922, in index_expr
return self._inner.index_expr(self._simplify(index), dtype)
File "", line 188, in index_expr
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/common.py", line 2296, in _default
value = getattr(self.parent_handler, name)(*args, **kwargs) # type: ignore[has-type]
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/mps.py", line 143, in index_expr
idx_str = V.kernel.index_to_str(V.kernel.prepare_indexing(expr))
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 815, in prepare_indexing
return self.codegen_indexing(simp_index)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 843, in codegen_indexing
self.range_tree_nodes[sym].codegen() # type: ignore[index]
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/simd.py", line 296, in _codegen
V.kernel.codegen_iteration_ranges_entry(self)
File ".../venv/lib/python3.11/site-packages/torch/_inductor/codegen/mps.py", line 550, in codegen_iteration_ranges_entry
if not entry.is_reduction or not self.multistage_reduction:
File ".../venv/lib/python3.11/site-packages/sympy/core/relational.py", line 519, in bool
raise TypeError(
torch._inductor.exc.InductorError: TypeError: cannot determine truth value of Relational: s0 > 1024

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

[timbo1975]

Just wanted to update and say that now I've used ComfyUI for ACE-Step I get no crashes and less memory usage (possibly because they always use the latest pytorch?). I have found generation quicker, RAM usage lower and I have to say the quality of the music even only using sometimes 7-10 steps is brilliant (I've not played fully with Stable Audio but so far ACE-Step seems to produce better songs at the sacrifice of audio quality but both are amazing developments).

I'm not including the voice yet (using the [inst] tag) I'm more interested in a good beat/tune then layer on vocals (ha sounds like I even know what I'm doing - I don't!)