AmosLewis

Args: ./../torch-mlir/build/bin/torch-mlir-opt -pass-pipeline=torch-backend-to-linalg-on-tensors-backend-pipeline --debug ./shark_tmp/_lambda.mlir 
Load new dialect in Context builtin
ImplicitTypeIDRegistry::lookupOrInsert(mlir::SubElementTypeInterface)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::ShapedType)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::MemRefLayoutAttrInterface)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::SubElementAttrInterface)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::ElementsAttr)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::TypedAttr)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::SymbolOpInterface)
ImplicitTypeIDRegistry::lookupOrInsert(mlir::OpAsmOpInterface)

AmosLewis

python tank/bloom_model.py
with torch_mlir_lockstep_tensor.py

......
......

Mismatched elements: 131059 / 131072 (100%)
Max absolute difference: 2.3001497
Max relative difference: 2833.5105
 x: array([[[-0.013242, -0.013242, -0.013242, ..., -0.013242, -0.013242,

AmosLewis

➜  SHARK git:(bloom) ✗ python tank/bloom_model.py
Some weights of BloomForSequenceClassification were not initialized from the model checkpoint at bigscience/bloom-560m and are newly initialized: ['score.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
/home/chi/src/ubuntu20/shark/SHARK/shark.venv/lib/python3.10/site-packages/torch/jit/_check.py:181: UserWarning: The TorchScript type system doesn't support instance-level annotations on empty non-base types in `__init__`. Instead, either 1) use a type annotation in the class body, or 2) wrap the type in `torch.jit.Attribute`.
  warnings.warn("The TorchScript type system doesn't support "
/home/chi/src/ubuntu20/shark/SHARK/shark.venv/lib/python3.10/site-packages/torch/jit/_trace.py:744: UserWarning: The input to trace is already a ScriptModule, tracing it is a no-op. Returning the object as is.
  warnings.warn(
Target triple found:x86_64-linux-gnu
tensor([[  7.2041, -17.0263]], grad_fn=<Inde

AmosLewis

https://github.com/llvm/torch-mlir/blob/6c1dea1c0ff22efb7119f6453655b8b38b52e506/lib/Conversion/TorchToLinalg/Reduction.cpp#L409
   ->
        result = convertScalarToDtype(rewriter, loc, result, mlir::IntegerType::get(op->getContext(), 64));


➜  SHARK git:(bloom) ✗ python tank/bloom_model.py
Some weights of BloomForSequenceClassification were not initialized from the model checkpoint at bigscience/bloom-560m and are newly initialized: ['score.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
/home/chi/src/ubuntu20/shark/SHARK/shark.venv/lib/python3.10/site-packages/torch/jit/_check.py:181: UserWarning: The TorchScript type system doesn't support instance-level annotations on empty non-base types in `__init__`. Instead, either 1) use a type annotation in the class body, or 2) wrap the type in `torch.jit.Attribute`.
  warnings.warn("The TorchScript type system doesn't support "

AmosLewis

import torch
from transformers import AutoModelForSequenceClassification

class HuggingFaceLanguage(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.model = AutoModelForSequenceClassification.from_pretrained(
            "bigscience/bloom-560m",  # The pretrained model.
            num_labels=2,  # The number of output labels--2 for binary classification.
            output_attentions=False,  # Whether the model returns attentions weights.

AmosLewis

func.func @torch.aten.Int.Tensor$zero_rank(%arg0: !torch.vtensor<[],ui8>) -> !torch.int {
  %0 = torch.aten.Int.Tensor %arg0 : !torch.vtensor<[],ui8> -> !torch.int
  return %0 : !torch.int
}

AmosLewis

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

from torch.fx.experimental.proxy_tensor import make_fx
from torch._decomp import get_decompositions
import tempfile
import torch_mlir


def prepare_sentence_tokens(hf_model: str, sentence: str):

AmosLewis

# from torch.fx import symbolic_trace
# # Symbolic tracing frontend - captures the semantics of the module
# symbolic_traced : torch.fx.GraphModule = symbolic_trace(fx_g)
# # High-level intermediate representation (IR) - Graph representation
# print(symbolic_traced.graph)

graph():
    %arg0_1 : [#users=2] = placeholder[target=arg0_1]
    %view : [#users=1] = call_function[target=torch.ops.aten.view](args = (%arg0_1, [-1, 128]), kwargs = {})
    %_param_constant0 : [#users=1] = get_attr[target=_param_constant0]

AmosLewis

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

from torch.fx.experimental.proxy_tensor import make_fx
from torch._decomp import get_decompositions
import tempfile
import torch_mlir

def prepare_sentence_tokens(hf_model: str, sentence: str):
    tokenizer = AutoTokenizer.from_pretrained(hf_model)

AmosLewis

#loc0 = loc(unknown)
module attributes {torch.debug_module_name = "_lambda"} {
  func.func @forward(%arg0: !torch.vtensor<[1,5],si64> loc(unknown)) -> !torch.vtensor<[1,5,50257],f32> {
    %int5 = torch.constant.int 5 loc(#loc1)
    %int1 = torch.constant.int 1 loc(#loc2)
    %true = torch.constant.bool true loc(#loc3)
    %float0.000000e00 = torch.constant.float 0.000000e+00 loc(#loc4)
    %0 = torch.vtensor.literal(dense_resource<__elided__> : tensor<768xf32>) : !torch.vtensor<[768],f32> loc(#loc0)
    %1 = torch.vtensor.literal(dense_resource<__elided__> : tensor<768xf32>) : !torch.vtensor<[768],f32> loc(#loc0)
    %2 = torch.vtensor.literal(dense_resource<__elided__> : tensor<3072x768xf32>) : !torch.vtensor<[3072,768],f32> loc(#loc0)