Files
pytorch/torch/csrc/jit/frontend/function_schema_parser.cpp
Han Qi 61d6c43864 Make debug_pkl smaller by only emitting unique traces. (#73368)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73368

debug_pkl file inside of pytorch's .pt file consists of a list of SourceRanges. Each SourceRange points to a Source which is a stack track, filename, and start, end numbers. Those are emitted in debug_pkl file as strings.
Since many SourceRange shares the same source, the string for trace can be deduped.
The newer format saves a set of unique traces in a tuple, then each SourceRange will save the offset of it's trace w.r.t. position in that tuple. (i.e. manually applying dictionary compression).
The above helps with smaller file size. On loading, if we copy each trace to Source as string the runtime memory would still blowup.
To mitigate this, we use SourceView directly instead of source which will take the reference of string inside of Deserializer and make that into string_view. This is safe because Deserializer is hold by Unpickler by shared_ptr, and Unpickler is also hold by shared_ptr by another Source object. That Source object will be alive during the model construction.

Test Plan:
unit test

Took original file (312271638_930.predictor.disagg.local); loaded with `torch.jit.load` save again with `torch.jit.save`. Unzip both, look at contents:
```
[qihan@devvm5585.vll0 ~]$ du archive -h
4.0K    archive/xl_model_weights
3.7M    archive/extra
8.0K    archive/code/__torch__/caffe2/torch/fb/model_transform/splitting
8.0K    archive/code/__torch__/caffe2/torch/fb/model_transform
8.0K    archive/code/__torch__/caffe2/torch/fb
8.0K    archive/code/__torch__/caffe2/torch
8.0K    archive/code/__torch__/caffe2
20M     archive/code/__torch__/torch/fx/graph_module
20M     archive/code/__torch__/torch/fx
8.0K    archive/code/__torch__/torch/classes
20M     archive/code/__torch__/torch
20M     archive/code/__torch__
20M     archive/code
2.7M    archive/constants
35M     archive
[qihan@devvm5585.vll0 ~]$ du resaved -h
4.0K    resaved/extra
8.0K    resaved/code/__torch__/caffe2/torch/fb/model_transform/splitting
8.0K    resaved/code/__torch__/caffe2/torch/fb/model_transform
8.0K    resaved/code/__torch__/caffe2/torch/fb
8.0K    resaved/code/__torch__/caffe2/torch
8.0K    resaved/code/__torch__/caffe2
1.3M    resaved/code/__torch__/torch/fx/graph_module
1.3M    resaved/code/__torch__/torch/fx
8.0K    resaved/code/__torch__/torch/classes
1.4M    resaved/code/__torch__/torch
1.4M    resaved/code/__torch__
1.4M    resaved/code
2.7M    resaved/constants
13M     resaved
[qihan@devvm5585.vll0 ~]$
```

Reviewed By: gmagogsfm

Differential Revision: D34455360

fbshipit-source-id: 8cc716f9bba7183746b1b4ecc33a2de34ac503b9
(cherry picked from commit f1a04730fc9ac8fdab6c8e4c44cb5529e42090e4)
2022-03-02 08:37:08 +00:00

382 lines
12 KiB
C++

#include <torch/csrc/jit/frontend/function_schema_parser.h>
#include <ATen/core/Reduction.h>
#include <ATen/core/type_factory.h>
#include <c10/util/Optional.h>
#include <c10/util/string_utils.h>
#include <torch/csrc/jit/frontend/lexer.h>
#include <torch/csrc/jit/frontend/parse_string_literal.h>
#include <torch/csrc/jit/frontend/schema_type_parser.h>
#include <functional>
#include <memory>
#include <vector>
using at::TypeKind;
using c10::Argument;
using c10::either;
using c10::FunctionSchema;
using c10::IValue;
using c10::ListType;
using c10::make_left;
using c10::make_right;
using c10::OperatorName;
using c10::OptionalType;
namespace torch {
namespace jit {
namespace {
struct SchemaParser {
explicit SchemaParser(const std::string& str)
: L(std::make_shared<Source>(
c10::string_view(str),
c10::nullopt,
0,
nullptr,
Source::DONT_COPY)),
type_parser(L, /*parse_complete_tensor_types*/ false) {}
either<OperatorName, FunctionSchema> parseDeclaration() {
OperatorName name = parseName();
// If there is no parentheses coming, then this is just the operator name
// without an argument list
if (L.cur().kind != '(') {
return make_left<OperatorName, FunctionSchema>(std::move(name));
}
std::vector<Argument> arguments;
std::vector<Argument> returns;
bool kwarg_only = false;
bool is_vararg = false;
bool is_varret = false;
size_t idx = 0;
parseList('(', ',', ')', [&] {
if (is_vararg)
throw ErrorReport(L.cur())
<< "... must be the last element of the argument list";
if (L.nextIf('*')) {
kwarg_only = true;
} else if (L.nextIf(TK_DOTS)) {
is_vararg = true;
} else {
arguments.push_back(parseArgument(
idx++, /*is_return=*/false, /*kwarg_only=*/kwarg_only));
}
});
// check if all arguments are not-default for vararg schemas
if (is_vararg) {
for (const auto& arg : arguments) {
if (arg.default_value().has_value()) {
throw ErrorReport(L.cur())
<< "schemas with vararg (...) can't have default value args";
}
}
}
idx = 0;
L.expect(TK_ARROW);
if (L.nextIf(TK_DOTS)) {
is_varret = true;
} else if (L.cur().kind == '(') {
parseList('(', ',', ')', [&] {
if (is_varret) {
throw ErrorReport(L.cur())
<< "... must be the last element of the return list";
}
if (L.nextIf(TK_DOTS)) {
is_varret = true;
} else {
returns.push_back(
parseArgument(idx++, /*is_return=*/true, /*kwarg_only=*/false));
}
});
} else {
returns.push_back(
parseArgument(0, /*is_return=*/true, /*kwarg_only=*/false));
}
return make_right<OperatorName, FunctionSchema>(
std::move(name.name),
std::move(name.overload_name),
std::move(arguments),
std::move(returns),
is_vararg,
is_varret);
}
c10::OperatorName parseName() {
std::string name = L.expect(TK_IDENT).text();
if (L.nextIf(':')) {
L.expect(':');
name = name + "::" + L.expect(TK_IDENT).text();
}
std::string overload_name = "";
if (L.nextIf('.')) {
overload_name = L.expect(TK_IDENT).text();
}
// default is used as an attribute on the `OpOverloadPacket`
// (obtained using `torch.ops.aten.foo`) to get the operator
// overload with overload name as an empty string
// and so shouldn't be used as an overload name
// also disallow dunder attribute names to be overload names
bool is_a_valid_overload_name =
!((overload_name == "default") || (overload_name.rfind("__", 0) == 0));
TORCH_CHECK(
is_a_valid_overload_name,
overload_name,
" is not a legal overload name for aten operators");
return {name, overload_name};
}
std::vector<either<OperatorName, FunctionSchema>> parseDeclarations() {
std::vector<either<OperatorName, FunctionSchema>> results;
do {
results.push_back(parseDeclaration());
} while (L.nextIf(TK_NEWLINE));
L.expect(TK_EOF);
return results;
}
either<OperatorName, FunctionSchema> parseExactlyOneDeclaration() {
auto result = parseDeclaration();
L.nextIf(TK_NEWLINE);
L.expect(TK_EOF);
return result;
}
Argument parseArgument(size_t idx, bool is_return, bool kwarg_only) {
auto p = type_parser.parseType();
auto type = std::move(p.first);
auto alias_info = std::move(p.second);
c10::optional<int32_t> N;
c10::optional<IValue> default_value;
c10::optional<std::string> alias_set;
std::string name;
if (L.nextIf('[')) {
// note: an array with a size hint can only occur at the Argument level
type = ListType::create(std::move(type));
N = c10::stoll(L.expect(TK_NUMBER).text());
L.expect(']');
auto container = type_parser.parseAliasAnnotation();
if (container && alias_info) {
container->addContainedType(std::move(*alias_info));
}
alias_info = std::move(container);
if (L.nextIf('?')) {
type = c10::TypeFactory::create<c10::OptionalType>(std::move(type));
}
}
if (is_return) {
// optionally field names in return values
if (L.cur().kind == TK_IDENT) {
name = L.next().text();
} else {
name = "";
}
} else {
name = L.expect(TK_IDENT).text();
if (L.nextIf('=')) {
default_value = parseDefaultValue(*type, type->kind(), N);
}
}
return Argument(
std::move(name),
std::move(type),
N,
std::move(default_value),
!is_return && kwarg_only,
std::move(alias_info));
}
IValue parseSingleConstant(const c10::Type& type, TypeKind kind) {
if (kind == c10::TypeKind::DynamicType) {
return parseSingleConstant(
type, type.expectRef<c10::DynamicType>().dynamicKind());
}
switch (L.cur().kind) {
case TK_TRUE:
L.next();
return true;
case TK_FALSE:
L.next();
return false;
case TK_NONE:
L.next();
return IValue();
case TK_STRINGLITERAL: {
auto token = L.next();
return parseStringLiteral(token.range, token.text());
}
case TK_IDENT: {
auto tok = L.next();
auto text = tok.text();
if ("float" == text) {
return static_cast<int64_t>(at::kFloat);
} else if ("complex" == text) {
return static_cast<int64_t>(at::kComplexFloat);
} else if ("long" == text) {
return static_cast<int64_t>(at::kLong);
} else if ("strided" == text) {
return static_cast<int64_t>(at::kStrided);
} else if ("Mean" == text) {
return static_cast<int64_t>(at::Reduction::Mean);
} else if ("contiguous_format" == text) {
return static_cast<int64_t>(c10::MemoryFormat::Contiguous);
} else {
throw ErrorReport(L.cur().range) << "invalid numeric default value";
}
}
default:
std::string n;
if (L.nextIf('-'))
n = "-" + L.expect(TK_NUMBER).text();
else
n = L.expect(TK_NUMBER).text();
if (kind == TypeKind::ComplexType || n.find('j') != std::string::npos) {
auto imag = c10::stod(n.substr(0, n.size() - 1));
return c10::complex<double>(0, imag);
} else if (
kind == TypeKind::FloatType || n.find('.') != std::string::npos ||
n.find('e') != std::string::npos) {
return c10::stod(n);
} else {
int64_t v = c10::stoll(n);
return v;
}
}
}
IValue convertToList(
const c10::Type& type,
TypeKind kind,
const SourceRange& range,
const std::vector<IValue>& vs) {
switch (kind) {
case TypeKind::ComplexType:
return fmap(vs, [](const IValue& v) { return v.toComplexDouble(); });
case TypeKind::FloatType:
return fmap(vs, [](const IValue& v) { return v.toDouble(); });
case TypeKind::IntType:
return fmap(vs, [](const IValue& v) { return v.toInt(); });
case TypeKind::BoolType:
return fmap(vs, [](const IValue& v) { return v.toBool(); });
case TypeKind::DynamicType:
return convertToList(
type, type.expectRef<c10::DynamicType>().dynamicKind(), range, vs);
default:
throw ErrorReport(range)
<< "lists are only supported for float, int and complex types";
}
}
IValue parseConstantList(const c10::Type& type, TypeKind kind) {
auto tok = L.expect('[');
std::vector<IValue> vs;
if (L.cur().kind != ']') {
do {
vs.push_back(parseSingleConstant(type, kind));
} while (L.nextIf(','));
}
L.expect(']');
return convertToList(type, kind, tok.range, vs);
}
IValue parseTensorDefault(const SourceRange& range) {
L.expect(TK_NONE);
return IValue();
}
IValue parseDefaultValue(
const c10::Type& arg_type,
TypeKind kind,
c10::optional<int32_t> arg_N) {
auto range = L.cur().range;
switch (kind) {
case TypeKind::TensorType:
case TypeKind::GeneratorType:
case TypeKind::QuantizerType: {
return parseTensorDefault(range);
} break;
case TypeKind::StringType:
case TypeKind::OptionalType:
case TypeKind::NumberType:
case TypeKind::IntType:
case TypeKind::BoolType:
case TypeKind::FloatType:
case TypeKind::ComplexType:
return parseSingleConstant(arg_type, kind);
break;
case TypeKind::DeviceObjType: {
auto device_text =
parseStringLiteral(range, L.expect(TK_STRINGLITERAL).text());
return c10::Device(device_text);
break;
}
case TypeKind::ListType: {
auto elem_type = arg_type.containedType(0);
if (L.cur().kind == TK_IDENT) {
return parseTensorDefault(range);
} else if (arg_N && L.cur().kind != '[') {
IValue v = parseSingleConstant(*elem_type, elem_type->kind());
std::vector<IValue> repeated(*arg_N, v);
return convertToList(*elem_type, elem_type->kind(), range, repeated);
} else {
return parseConstantList(*elem_type, elem_type->kind());
}
} break;
case TypeKind::DynamicType:
return parseDefaultValue(
arg_type,
arg_type.expectRef<c10::DynamicType>().dynamicKind(),
arg_N);
default:
throw ErrorReport(range) << "unexpected type, file a bug report";
}
return IValue(); // silence warnings
}
void parseList(
int begin,
int sep,
int end,
c10::function_ref<void()> callback) {
auto r = L.cur().range;
if (begin != TK_NOTHING)
L.expect(begin);
if (L.cur().kind != end) {
do {
callback();
} while (L.nextIf(sep));
}
if (end != TK_NOTHING)
L.expect(end);
}
Lexer L;
SchemaTypeParser type_parser;
};
} // namespace
C10_EXPORT either<OperatorName, FunctionSchema> parseSchemaOrName(
const std::string& schemaOrName) {
return SchemaParser(schemaOrName).parseExactlyOneDeclaration();
}
C10_EXPORT FunctionSchema parseSchema(const std::string& schema) {
auto parsed = parseSchemaOrName(schema);
TORCH_CHECK(
parsed.is_right(),
"Tried to parse a function schema but only the operator name was given");
return std::move(parsed.right());
}
C10_EXPORT OperatorName parseName(const std::string& name) {
auto parsed = parseSchemaOrName(name);
TORCH_CHECK(
parsed.is_left(),
"Tried to parse an operator name but function schema was given");
return std::move(parsed.left());
}
} // namespace jit
} // namespace torch