jroesch commented on code in PR #11199: URL: https://github.com/apache/tvm/pull/11199#discussion_r863737379
########## python/tvm/relay/transform/flexible_shape.py: ########## @@ -0,0 +1,273 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +# pylint: disable=invalid-name, dangerous-default-value +"""Relay functions for wrapping a module with flexible shape dispatch.""" +from tvm import relay + + +def override_shape(tensor_type, dim, value): + """Change a value in a tensor shape.""" + new_dims = list(tensor_type.shape) + new_dims[dim] = value + return relay.TensorType(new_dims, tensor_type.dtype) + + +def specialize_body(mod, function, dim, value, input_indices=[0], affects_output=True): + """Create a subgraph to handle specific input shapes""" + # Iterate through specified inputs and construct specialized shapes for each. + new_params = list(function.params) + data_binding = {} + dyn_data_array = [] + for inp in input_indices: + data = function.params[inp] + flex_ty = override_shape(data.type_annotation, dim, value) + dyn_data = relay.Var(data.name_hint, type_annotation=flex_ty) + new_params[inp] = dyn_data + data_binding[data] = dyn_data + dyn_data_array.append(dyn_data) + + # Create a new function body for the modified shapes. + new_body = relay.expr.bind(function.body, data_binding) + # Only change the output shape if the input shape affects it. + if affects_output: + new_ret_ty = override_shape(function.ret_type, dim, value) + else: + new_ret_ty = function.ret_type + gvar = relay.GlobalVar("main_" + str(value)) + # Add the new function to the main IRModule. + mod[gvar] = relay.Function( + new_params, new_body, new_ret_ty, function.type_params, function.attrs + ) + return gvar, [d.type_annotation for d in dyn_data_array] + + +def flexible_dispatch( + mod, dim=0, buckets=[1], auto_pad=False, pad_value=0, input_indices=[0], affects_output=True +): + """ + Enable inference of multiple shaped inputs in one module. + + This transformation adds a handler around a module that + checks input shapes and dispatches to a subgraph specialized + to handle the specific shapes of that input. If no exactly matching + subgraph is available, the input will be run using full dynamism. + For best performance, specify all the sizes the module will + be likely to see using the buckets argument. + + Parameters + ---------- + dim: int + The dimension of the input that should be made flexible. This will + most often be used for the batch dimension. + buckets: list[int] + The sizes of the input dimension that should be explicitly handled. + Each value in buckets will have a corresponding subgraph constructed to + handle it. + auto_pad: Optional[bool] + If True, then padding will be inserted to values that don't match one of + the provided buckets. + pad_value: Optional[float] + When auto_pad is true, padding will be done with this value. + input_indices: Optional[List[int]] + Which inputs should be dispatched dynamically, provided by index. All inputs + must share the same dynamic axis. + affects_output: Optional[bool] + Whether the change in input shape has a corresponding effect on the output shape. + Batching for example effects both the input and output whereas changing sequence + length in an NLP model typically does not. + + Returns + ------- + mod : IRModule + The new module wrapped with a flexible shape dispatch handler. + """ + main_fn = mod["main"] + + # Extract all input data and create a new dynamic variable for each. + data = [] + dyn_data = [] + for i in input_indices: + data.append(main_fn.params[i]) + dyn_shape = override_shape(data[i].type_annotation, dim, relay.Any()) + dyn_data.append(relay.Var(data[i].name_hint, type_annotation=dyn_shape)) + + # Extract the dynamic shape value from one of the inputs. + rt_sh = relay.op.shape_of(dyn_data[0]) + flex_value = relay.op.take(rt_sh, relay.const(dim)) + + if_exprs = [] + + for i, bucket in enumerate(buckets): + input_data = dyn_data + check_dim = flex_value + + # Apply automatic padding if specified. + if auto_pad: + input_data = [] + # Construct padding expression for inputs. + for j, inp in enumerate(dyn_data): + pad_width = relay.const(bucket) - flex_value + rank = len(data[j].type_annotation.shape) + pads = relay.zeros([rank, 2], "int32") + pads = relay.scatter_nd(pads, relay.const([dim, 1]), pad_width) + padded_value = relay.nn.pad(inp, pads, pad_value) + + # Determine if this is the proper bucket to pad to. Do this by checking if the + # input shape is between this bucket and the previous. + if i == 0: + padded_value = relay.If( + relay.op.less_equal(flex_value, relay.const(bucket)), padded_value, inp + ) + else: + padded_value = relay.If( + relay.op.logical_and( + relay.op.less_equal(flex_value, relay.const(bucket)), + relay.op.greater(flex_value, relay.const(buckets[i - 1])), + ), + padded_value, + inp, + ) + # Update input value and test dimension to reflect possible padding. + input_data.append(padded_value) + # Grab the new possibly padded shape for checking bucket size. + check_dim = relay.op.take(relay.op.shape_of(input_data[0]), relay.const(dim)) + + # Create a specialized subgraph for the current bucket. + spec_call, spec_ty = specialize_body( + mod, main_fn, dim, bucket, input_indices=input_indices, affects_output=affects_output + ) + # Apply hard casting to shape to create statically typed graphs. + spec_data = [] + for j, inp in enumerate(input_data): + spec_data.append(relay.op.reshape(inp, spec_ty[j].shape)) + + # Create a dispatch statement for the current specialized graph. + call_args = list(main_fn.params) + for j, inp in enumerate(input_indices): + call_args[inp] = spec_data[j] + new_call = spec_call(*call_args) + + # Remove meaningless padded outputs if applicable. + if auto_pad and affects_output: + new_call = relay.take( + new_call, + relay.arange(start=relay.const(0), stop=flex_value, dtype="int32"), + axis=dim, + ) + + # Add this new case to the dispatch handler. + if_exprs.append((relay.op.equal(check_dim, relay.const(bucket)), new_call)) + + # Create a subgraph to handle all other shapes. + default_dyn_call, _ = specialize_body( + mod, main_fn, dim, relay.Any(), input_indices=input_indices, affects_output=affects_output + ) + call_args = list(main_fn.params) + for j, inp in enumerate(input_indices): + call_args[inp] = dyn_data[j] + new_body = default_dyn_call(*call_args) + + # Create an If chain to dispatch shapes to the appropriate specialized subgraph. + for cond, true_branch in if_exprs: + new_body = relay.If(cond, true_branch, new_body) + + # Assign new parameters to the function. + new_params = list(main_fn.params) + for j, inp in enumerate(input_indices): + new_params[inp] = dyn_data[j] + + # Update the output shape to be dynamic if needed. + if affects_output: + dyn_ret_type = override_shape(main_fn.ret_type, dim, relay.Any()) + else: + dyn_ret_type = main_fn.ret_type + + # Assign the handler as the new entrypoint in the module. + new_main = relay.Function( + new_params, new_body, dyn_ret_type, main_fn.type_params, main_fn.attrs + ) + mod["main"] = new_main + return mod + + +class FlexibleShapeDispatch(object): + """Enable inference of multiple shaped inputs in one module. + + This transformation adds a handler around a module that + checks input shapes and dispatches to a subgraph specialized + to handle the specific shapes of that input. If no exactly matching + subgraph is available, the input will be run using full dynamism. + For best performance, specify all the sizes the module will + be likely to see using the buckets argument. + + Parameters + ---------- + dim: int + The dimension of the input that should be made flexible. This will + most often be used for the batch dimension. + buckets: list[int] + The sizes of the input dimension that should be explicitly handled. + Each value in buckets will have a corresponding subgraph constructed to + handle it. + auto_pad: Optional[bool] + If True, then padding will be inserted to values that don't match one of + the provided buckets. + pad_value: Optional[float] + When auto_pad is true, padding will be done with this value. + input_indices: Optional[List[int]] + Which inputs should be dispatched dynamically, provided by index. All inputs + must share the same dynamic axis. + affects_output: Optional[bool] + Whether the change in input shape has a corresponding effect on the output shape. + Batching for example effects both the input and output whereas changing sequence + length in an NLP model typically does not. + + Returns + ------- + ret : FlexibleShapeDispatch + A pass that can be applied to a module to add flexible shape handling. + """ + + def __init__( + self, + dim=0, + buckets=[1], Review Comment: same as above. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
