def test_debug_graph_runtime():
if not tvm.testing.device_enabled("llvm"):
print("Skip because llvm is not enabled")
return
mod, params = relay.testing.synthetic.get_workload()
with relay.build_config(opt_level=3):
complied_graph_lib = relay.build_module.build(mod, "llvm", params=params)
data = np.random.uniform(-1, 1, size=input_shape(mod)).astype("float32")
# raw api
ctx = tvm.cpu()
try:
gmod = complied_graph_lib["debug_create"]("default", ctx)
except:
print("Skip because debug graph_runtime not enabled")
return
set_input = gmod["set_input"]
run = gmod["run"]
get_output = gmod["get_output"]
set_input("data", tvm.nd.array(data))
run()
out = get_output(0).asnumpy()
tvm.testing.assert_allclose(out, verify(data), atol=1e-5)
# debug graph runtime wrapper
debug_g_mod = debug_runtime.GraphModuleDebug(
complied_graph_lib["debug_create"]("default", ctx),
[ctx],
complied_graph_lib.get_json(),
None,
)
debug_g_mod.set_input("data", data)
debug_g_mod.run()
out = debug_g_mod.get_output(0).asnumpy()
tvm.testing.assert_allclose(out, verify(data), atol=1e-5)
def eval_cpu(args, model_inst, compiled_model, samples):
lowered = model_inst.lower_model(compiled_model)
if args.use_debug_runtime:
graph_mod = debug_runtime.GraphModuleDebug(
lowered.module["debug_create"]("default", tvm.cpu(0)),
[tvm.cpu(0)], lowered.get_json(),
dump_root=f'{util.get_repo_root()}/debug/cpu')
print('graph mod', graph_mod)
else:
graph_mod = tvm.contrib.graph_runtime.GraphModule(lowered["default"](tvm.cpu(0)))
print('graph mod', lowered, graph_mod)
graph_mod.set_input(**lowered.params)
results = []
for i, sample in enumerate(samples):
inputs = adapt_all_inputs(
model_inst, sample,
compiled_model.ir_mod[compiled_model.entry_point])
for key in inputs:
if key in sample.inputs:
graph_mod.set_input(key, inputs[key])
graph_mod.run()
results.append(adapt_all_outputs(model_inst, graph_mod, sample))
_LOG.info('got prediction: %r', results[-1])
return results