[GitHub] [tvm] elvin-n opened a new issue, #14936: [Bug] [Unity] [Metascheduler] cannot tune relax linear/matmul with M > 1 for cuda

["elvin-n (via GitHub)" <gi...@apache.org>]

elvin-n opened a new issue, #14936:
URL: https://github.com/apache/tvm/issues/14936

   Unable to tune linear/matmul having M value bigger than 1.
   The error message looks the same as in Issue#14935, but below is a definition through relax.linear
   If reasons the same, probably make sense to close this one as "same as" status
   
   ``` python
   import tvm
   from tvm import meta_schedule as ms
   from tvm.relay.backend import Executor
   from tvm import relax
   from tvm.relax.testing import nn
   
   # -------- Func definition
   class Linear(nn.Module):
       def __init__(self, in_features, out_features, dtype: str, bias=False):
           self.in_features = in_features
           self.out_features = out_features
           self.weight = nn.Parameter(
               (out_features, in_features), dtype=dtype, name="linear_weight"
           )
           if bias:
               self.bias = nn.Parameter((out_features,), dtype=dtype, name="linear_bias")
           else:
               self.bias = None
   
       def forward(self, input: relax.Expr) -> relax.Var:
           return nn.emit(relax.op.linear(input, self.weight, self.bias))
   
   bb = relax.BlockBuilder()
   seq_len = 4
   with bb.function("func1"):
       model = Linear(2048, 768, "float16")
       input = nn.Placeholder((seq_len, 2048), dtype="float16", name="input")
       with bb.dataflow():
           res = model(input)
           params = [
               input,
           ] + model.parameters()
           gv = bb.emit_output((res,))
       bb.emit_func_output(gv, params)
   
   mod = bb.get()
   gv = mod.get_global_var("func1")
   bb.update_func(gv, mod[gv].with_attr("func1", 1))
   
   mod = relax.pipeline.get_pipeline()(mod)
   mod = relax.transform.LiftTransformParams()(mod)
   
   mod = tvm.tir.transform.ForceNarrowIndexToInt32()(mod)
   
   # ------ Metascheduler starts here
   database = None
   
   strategy_name = "evolutionary"
   name = f"relax_linear_{seq_len}_2048_2048_768"
   work_dir = f"./{name}/"
   module_equality_name = "ignore-ndarray"
   
   target = tvm.target.Target("nvidia/geforce-rtx-2060", host="llvm")
   executor = Executor("graph")
   mod = mod.with_attr("executor", executor)
   ndk_builder = ms.builder.LocalBuilder(timeout_sec=60)
   evaluator_config=ms.runner.EvaluatorConfig(
       number=3,
       repeat=1,
       min_repeat_ms=100,
       enable_cpu_cache_flush=False,
   )
   ms_rpc_runner = ms.runner.LocalRunner(evaluator_config=evaluator_config,
               alloc_repeat=1,
           )
   ms.relax_integration.tune_relax(
       mod=mod,
       target=target,
       params={},
       work_dir=work_dir,
       max_trials_global=1024,
       strategy=strategy_name,
       builder=ndk_builder,
       runner=ms_rpc_runner,
       module_equality=module_equality_name,
   )
   ```


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org