[fix][inductor] fix master triton acc-check bug-pytorch-AtomGit

Mtorch_npu/_inductor/codegen/ir_fx.py+3-3

	@@ -318,9 +318,9 @@ def _patch_baseview_realize_hint(self):
318	else:	318	else:
319	return self.data.realize_hint()	319	return self.data.realize_hint()
320		320
321	-def _patch_mark_reuse(self, users):	321	+def _patch_mark_reuse(self, users, **kwargs):
322	if hasattr(self, 'traced_graph') and self.traced_graph is not None:	322	if hasattr(self, 'traced_graph') and self.traced_graph is not None:
323	- r = self.data.mark_reuse(users)	323	+ r = self.data.mark_reuse(users, **kwargs)
324	buffer = try_get_buffer(self)	324	buffer = try_get_buffer(self)
325	if not buffer:	325	if not buffer:
326	return r	326	return r
	@@ -340,7 +340,7 @@ def _patch_mark_reuse(self, users):
340	self._post_init_setattr("traced_graph", new_traced_graph)	340	self._post_init_setattr("traced_graph", new_traced_graph)
341	return r	341	return r
342	else:	342	else:
343	- return self.data.mark_reuse(users)	343	+ return self.data.mark_reuse(users, **kwargs)
344		344
345	@classmethod	345	@classmethod
346	def _patch_expandview_create(cls, x, new_size, traced_graph=None, node_name=None):	346	def _patch_expandview_create(cls, x, new_size, traced_graph=None, node_name=None):

Mtorch_npu/_inductor/codegen/scheduling.py+10-6

	@@ -159,6 +159,7 @@ class NPUTritonScheduling(TritonScheduling):
159	V.graph.inplaced_to_remove \|= kernel.inplaced_to_remove	159	V.graph.inplaced_to_remove \|= kernel.inplaced_to_remove
160		160
161	traced_graph_hash = None	161	traced_graph_hash = None
		162	+ kernel_graph_hash = None
162	if npu_config.dump_fx_graph:	163	if npu_config.dump_fx_graph:
163	if not npu_config.traced_fx_graph_cache:	164	if not npu_config.traced_fx_graph_cache:
164	npu_config.traced_fx_graph_cache = os.path.join(os.getenv("TORCHINDUCTOR_CACHE_DIR"),	165	npu_config.traced_fx_graph_cache = os.path.join(os.getenv("TORCHINDUCTOR_CACHE_DIR"),
	@@ -168,9 +169,11 @@ class NPUTritonScheduling(TritonScheduling):
168	if traced_graph is None:	169	if traced_graph is None:
169	log.warning("For nodes %s, could not gen fx graph while dump-graph.", nodes)	170	log.warning("For nodes %s, could not gen fx graph while dump-graph.", nodes)
170	else:	171	else:
171	- traced_graph_hash = code_hash(src_code + traced_graph.code)	172	+ fx_arg_shapes_str = str([getattr(a, "shape", a) for a in fx_args])
		173	+ traced_graph_hash = code_hash(src_code + traced_graph.code + fx_arg_shapes_str)
		174	+ kernel_graph_hash = code_hash(src_code + traced_graph.code)
172		175
173	- kernel_name, src_code = self.define_kernel(src_code, node_schedule, kernel, traced_graph_hash)	176	+ kernel_name, src_code = self.define_kernel(src_code, node_schedule, kernel, traced_graph_hash, kernel_graph_hash)
174		177
175	kernel.kernel_name = kernel_name	178	kernel.kernel_name = kernel_name
176	kernel.code_hash = code_hash(src_code)	179	kernel.code_hash = code_hash(src_code)
	@@ -366,10 +369,11 @@ class NPUTritonScheduling(TritonScheduling):
366	kernel_code_list.append((src_code, kernel, node_group))	369	kernel_code_list.append((src_code, kernel, node_group))
367	return kernel_code_list	370	return kernel_code_list
368		371
369	- def define_kernel(self, src_code, node_schedule, kernel, traced_graph_hash: str):	372	+ def define_kernel(self, src_code, node_schedule, kernel, traced_graph_hash: str, kernel_graph_hash: str = None):
370	wrapper = V.graph.wrapper_code	373	wrapper = V.graph.wrapper_code
371	- if (src_code, traced_graph_hash) in wrapper.src_to_kernel:	374	+ kernel_cache_key = (src_code, kernel_graph_hash)
372	- kernel_name = wrapper.src_to_kernel[(src_code, traced_graph_hash)]	375	+ if kernel_cache_key in wrapper.src_to_kernel:
		376	+ kernel_name = wrapper.src_to_kernel[kernel_cache_key]
373	if npu_config.dump_fx_graph:	377	if npu_config.dump_fx_graph:
374	src_code = src_code.replace(str(Placeholder.DESCRIPTIVE_NAME), kernel_name)	378	src_code = src_code.replace(str(Placeholder.DESCRIPTIVE_NAME), kernel_name)
375	subs_name = kernel_name if config.triton.unique_kernel_names else "triton_"	379	subs_name = kernel_name if config.triton.unique_kernel_names else "triton_"
	@@ -390,7 +394,7 @@ class NPUTritonScheduling(TritonScheduling):
390	["triton", kernel_category, fused_name, wrapper.next_kernel_suffix()]	394	["triton", kernel_category, fused_name, wrapper.next_kernel_suffix()]
391	)	395	)
392	# use the original src_code as the key	396	# use the original src_code as the key
393	- wrapper.src_to_kernel[(src_code, traced_graph_hash)] = kernel_name	397	+ wrapper.src_to_kernel[kernel_cache_key] = kernel_name
394	subs_name = kernel_name if config.triton.unique_kernel_names else "triton_"	398	subs_name = kernel_name if config.triton.unique_kernel_names else "triton_"
395		399
396	# DESCRIPTIVE_NAME is used for profiling purposes; it shows the full kernel name	400	# DESCRIPTIVE_NAME is used for profiling purposes; it shows the full kernel name

Mtorch_npu/_inductor/codegen/triton.py+6-0

	@@ -5,6 +5,7 @@ import functools
5	import itertools	5	import itertools
6	import math	6	import math
7	import operator	7	import operator
		8	+import os
8	import re	9	import re
9	from collections.abc import Callable, Iterable, Sequence	10	from collections.abc import Callable, Iterable, Sequence
10	from typing import Any, cast, Optional	11	from typing import Any, cast, Optional
	@@ -4261,6 +4262,11 @@ class NPUIndexTritonKernel(TritonKernel):
4261	return sympy_var	4262	return sympy_var
4262		4263
4263	if var.bounds.lower < 0: # type: ignore[operator]	4264	if var.bounds.lower < 0: # type: ignore[operator]
		4265	+ if (
		4266	+ os.environ.get("INDUCTOR_ASCEND_DUMP_FX_GRAPH")
		4267	+ or os.environ.get("INDUCTOR_ASCEND_CHECK_ACCURACY")
		4268	+ ):
		4269	+ wrap_neg = False
4264	if wrap_neg:	4270	if wrap_neg:
4265	stm = ops.add(var, ops.index_expr(size, torch.long))	4271	stm = ops.add(var, ops.index_expr(size, torch.long))
4266	# Mixed negative and non-negative	4272	# Mixed negative and non-negative

Mtorch_npu/_inductor/lowering.py+9-2

	@@ -72,8 +72,10 @@ from torch._inductor.scheduler import Scheduler
72		72
73	from .codegen.triton_utils import NPUKernelType	73	from .codegen.triton_utils import NPUKernelType
74	from .ir import IndexputTemplate, ScatterTemplate	74	from .ir import IndexputTemplate, ScatterTemplate
75	-from .lowering_common import (	75	+from .lowering_common import ( # noqa: F401 re-export for codegen/ir_fx.py
76	add_overload as _add_overload,	76	add_overload as _add_overload,
		77	+ create_fake_input,
		78	+ subtract_graph,
77	enable_full_lowering_fallback as enable_full_lowering_fallback_common,	79	enable_full_lowering_fallback as enable_full_lowering_fallback_common,
78	resolve_op_from_name,	80	resolve_op_from_name,
79	)	81	)
	@@ -83,10 +85,15 @@ from .config import inductor_indirect_memory_mode, log, is_ascend950, enable_ful
83	from .lowering_fallback_list import FALLBACK_LIST, NPU_EXTRA_FALLBACK_LIST	85	from .lowering_fallback_list import FALLBACK_LIST, NPU_EXTRA_FALLBACK_LIST
84		86
85	from . import config as npu_config	87	from . import config as npu_config
86	-from .lowering_fx import (	88	+from .lowering_fx import ( # noqa: F401 re-export for codegen/scheduling.py & ir_fx.py
87	fetch_graphs,	89	fetch_graphs,
88	merge_traced_graphs,	90	merge_traced_graphs,
89	node_id,	91	node_id,
		92	+ create_fx_from_snodes_by_traced_graph,
		93	+ create_compile_kwargs,
		94	+ generate_fx_graph_code,
		95	+ dump_fx_graph_code,
		96	+ snodes_to_fx,
90	)	97	)
91		98
92	def npu_make_fallback(op, layout_constraint=None, warn=True, override_decomp=False):	99	def npu_make_fallback(op, layout_constraint=None, warn=True, override_decomp=False):

Mtorch_npu/_inductor/lowering_common.py+15-2

	@@ -323,7 +323,8 @@ def fetch_graphs(
323	*,	323	*,
324	use_npu_meta: bool = False,	324	use_npu_meta: bool = False,
325	):	325	):
326	- if isinstance(inputs, (TensorBox, ir.StorageBox, ir.View, sympy.Symbol, ir.Constant, ir.ReinterpretView)):	326	+ if isinstance(inputs, (TensorBox, ir.StorageBox, ir.View, ir.ExpandView, ir.PermuteView, ir.SliceView,
		327	+ sympy.Symbol, ir.Constant, ir.ReinterpretView)):
327	inputs = [inputs]	328	inputs = [inputs]
328	input_graphs = []	329	input_graphs = []
329	for inp in inputs:	330	for inp in inputs:
	@@ -351,7 +352,19 @@ def fetch_graphs(
351	continue	352	continue
352	name = inp.get_name()	353	name = inp.get_name()
353	traced_graph = inp.get_traced_graph()	354	traced_graph = inp.get_traced_graph()
354	- if traced_graph is not None:	355	+ if (
		356	+ traced_graph is not None
		357	+ and not isinstance(inp, ir.ConcatKernel)
		358	+ and not (
		359	+ hasattr(inp, 'data')
		360	+ and isinstance(inp.data, ir.ConcatKernel)
		361	+ )
		362	+ and not (
		363	+ hasattr(inp, 'data')
		364	+ and hasattr(inp.data, 'data')
		365	+ and isinstance(inp.data.data, ir.ConcatKernel)
		366	+ )
		367	+ ):
355	input_graphs.append(traced_graph)	368	input_graphs.append(traced_graph)
356	continue	369	continue
357	traced_graph = TracedGraph()	370	traced_graph = TracedGraph()

Mtorch_npu/_inductor/lowering_fx.py+111-10

	@@ -445,11 +445,16 @@ def create_compile_kwargs(final_kernel, fx_call_args, fx_args):
445	fx_call_args[idx] = final_kernel.args.inplace_buffers[call_arg].other_names[-1]	445	fx_call_args[idx] = final_kernel.args.inplace_buffers[call_arg].other_names[-1]
446	fx_arg_shapes = [fx_arg.shape if isinstance(fx_arg, torch.Tensor) else ['ushape'] for fx_arg in fx_args]	446	fx_arg_shapes = [fx_arg.shape if isinstance(fx_arg, torch.Tensor) else ['ushape'] for fx_arg in fx_args]
447		447
		448	+ # add_numel 前移: 让 numel 表达式参与后续匹配, SymInt 标量输入有机会匹配到 kernel numel arg
		449	+ final_kernel.add_numel_to_call_args(final_kernel.kernel_name, kernel_call_args, arg_types)
		450	+ # add_numel 后 kernel_call_args 可能含 SymbolicCallArg (unhashable), 统一用 str 比较
		451	+ kernel_call_args_str = [str(arg) for arg in kernel_call_args]
		452	+
448	fx_call_args_no_dy = [arg for arg in fx_call_args if not (arg.startswith(("_", "u")))]	453	fx_call_args_no_dy = [arg for arg in fx_call_args if not (arg.startswith(("_", "u")))]
449		454
450	- if set(kernel_call_args) != set(fx_call_args_no_dy):	455	+ if set(kernel_call_args_str) != set(fx_call_args_no_dy):
451	# Handle NonOwningLayout alias	456	# Handle NonOwningLayout alias
452	- kernel_call_args_set = set(kernel_call_args)	457	+ kernel_call_args_set = set(kernel_call_args_str)
453	for idx, call_arg in enumerate(fx_call_args):	458	for idx, call_arg in enumerate(fx_call_args):
454	if call_arg in kernel_call_args_set:	459	if call_arg in kernel_call_args_set:
455	continue	460	continue
	@@ -462,9 +467,22 @@ def create_compile_kwargs(final_kernel, fx_call_args, fx_args):
462	break	467	break
463		468
464	fx_call_args_no_dy = [arg for arg in fx_call_args if not (arg.startswith(("_", "u")))]	469	fx_call_args_no_dy = [arg for arg in fx_call_args if not (arg.startswith(("_", "u")))]
465	- if set(kernel_call_args) != set(fx_call_args_no_dy):	470	+ kernel_call_args_str_set = set(kernel_call_args_str)
466	- return None	471	+ if kernel_call_args_str_set != set(fx_call_args_no_dy):
467	- final_kernel.add_numel_to_call_args(final_kernel.kernel_name, kernel_call_args, arg_types)	472	+ # 区分: 不匹配的是 "标量输入" (可恢复) 还是 "buffer" (应放弃)
		473	+ fx_only = set(fx_call_args_no_dy) - kernel_call_args_str_set
		474	+ scalar_only_mismatch = True
		475	+ for buf_name in fx_only:
		476	+ buf = V.graph.try_get_buffer(buf_name)
		477	+ if buf is not None:
		478	+ scalar_only_mismatch = False
		479	+ break
		480	+ if not scalar_only_mismatch:
		481	+ log.warning("[ACC-DEBUG] kernel=%s skipped: buffer mismatch (fx_only=%s)",
		482	+ final_kernel.kernel_name, fx_only)
		483	+ return None
		484	+ # scalar_only_mismatch: SymInt 标量输入没有对应 kernel arg, 用 -1 哨兵保留位置,
		485	+ # 运行期从输出 tensor shape 反推其值 (见 npu_compare.check_accuracy_triton)
468		486
469	kernel_call_args = [	487	kernel_call_args = [
470	map_operators_to_strings(str(item.inner_expr)) if isinstance(item, torch._inductor.codegen.wrapper.SymbolicCallArg)	488	map_operators_to_strings(str(item.inner_expr)) if isinstance(item, torch._inductor.codegen.wrapper.SymbolicCallArg)
	@@ -472,11 +490,75 @@ def create_compile_kwargs(final_kernel, fx_call_args, fx_args):
472	for item in kernel_call_args	490	for item in kernel_call_args
473	]	491	]
474	index_map = {str(element): idx for idx, element in enumerate(kernel_call_args)}	492	index_map = {str(element): idx for idx, element in enumerate(kernel_call_args)}
475	- call_args_mapping = [index_map[str(element)] for element in fx_call_args if str(element) in index_map]	493	+ # Prefer the runtime kernel argument for every exact match, including
		494	+ # symbolic scalar arguments such as a numel expression.
		495	+ call_args_mapping = []
		496	+ for i, element in enumerate(fx_call_args):
		497	+ element_str = str(element)
		498	+ if element_str in index_map:
		499	+ call_args_mapping.append(index_map[element_str])
		500	+ elif i < len(fx_args) and not isinstance(fx_args[i], torch.Tensor):
		501	+ # Only scalars absent from the kernel signature need a shape recipe.
		502	+ call_args_mapping.append(-1)
		503	+ else:
		504	+ return None
		505	+
		506	+ # 精确配方: 对每个 -1 标量输入, 编译期求解它等于哪个 kernel arg tensor 的第几维。
		507	+ # 序列化为具体 int (kernel_arg_idx, dim), 运行期 args[kidx].shape[dim] 精确反推,
		508	+ # 不依赖 "FX 标量位置 == 输出维度序号" 的假设 (比 shape[pos] 猜测更普适)。
		509	+ def _to_sympy_expr(val):
		510	+ if val is None:
		511	+ return None
		512	+ if isinstance(val, (int, sympy.Integer)):
		513	+ return sympy.sympify(val)
		514	+ if hasattr(val, "_sympy_"):
		515	+ try:
		516	+ return val._sympy_()
		517	+ except Exception:
		518	+ return None
		519	+ try:
		520	+ return sympy.sympify(val)
		521	+ except Exception:
		522	+ return None
		523	+
		524	+ scalar_resolved_from = {}
		525	+ for i, idx in enumerate(call_args_mapping):
		526	+ if idx != -1:
		527	+ continue
		528	+ scalar_expr = _to_sympy_expr(fx_args[i])
		529	+ if scalar_expr is None:
		530	+ return None
		531	+ found = None
		532	+ for j, fa in enumerate(fx_args):
		533	+ if not isinstance(fa, torch.Tensor):
		534	+ continue
		535	+ kidx = call_args_mapping[j]
		536	+ if kidx < 0:
		537	+ continue
		538	+ for d, dim_val in enumerate(fa.shape):
		539	+ dim_expr = _to_sympy_expr(dim_val)
		540	+ if dim_expr is None:
		541	+ continue
		542	+ if scalar_expr == dim_expr or sympy.simplify(scalar_expr - dim_expr) == 0:
		543	+ found = (kidx, d)
		544	+ break
		545	+ if found is not None:
		546	+ break
		547	+ if found is None:
		548	+ # 无法从任何 tensor 的 shape 反推该 SymInt 标量 → 跳过该 kernel 精度检查 (安全, 不猜)
		549	+ log.warning("[ACC-DEBUG] kernel=%s skipped: cannot resolve scalar at pos=%s from any tensor shape",
		550	+ final_kernel.kernel_name, i)
		551	+ return None
		552	+ scalar_resolved_from[i] = found
476	mismatch_indices_shapes = {}	553	mismatch_indices_shapes = {}
477		554
478	def is_dynamic_shape_dim(d):	555	def is_dynamic_shape_dim(d):
479	- if str(type(d)).find("torch.fx.experimental.symbolic_shapes") != -1:	556	+ type_str = str(type(d))
		557	+ if type_str.find("torch.fx.experimental.symbolic_shapes") != -1:
		558	+ return True
		559	+ if type_str.find("torch.SymInt") != -1:
		560	+ return True
		561	+ if type_str.find("sympy") != -1:
480	return True	562	return True
481	s = str(d).strip()	563	s = str(d).strip()
482	if s.startswith("u") or s.startswith("i") or s in ("-1", "?"):	564	if s.startswith("u") or s.startswith("i") or s in ("-1", "?"):
	@@ -487,15 +569,28 @@ def create_compile_kwargs(final_kernel, fx_call_args, fx_args):
487	return True	569	return True
488		570
489	for i in range(len(fx_call_args)):	571	for i in range(len(fx_call_args)):
490	- if any(is_dynamic_shape_dim(dim) for dim in fx_arg_shapes[i]):	572	+ if not isinstance(fx_args[i], torch.Tensor):
491	continue	573	continue
492	- mismatch_indices_shapes[i] = fx_arg_shapes[i]	574	+ reshape_shape = []
		575	+ dynamic_dim_count = 0
		576	+ for dim in fx_arg_shapes[i]:
		577	+ if is_dynamic_shape_dim(dim):
		578	+ dynamic_dim_count += 1
		579	+ reshape_shape.append(-1)
		580	+ else:
		581	+ reshape_shape.append(int(dim))
		582	+ # torch.reshape can infer one dynamic dimension from the runtime numel.
		583	+ if dynamic_dim_count > 1:
		584	+ continue
		585	+ mismatch_indices_shapes[i] = torch.Size(reshape_shape)
493		586
494	return {	587	return {
495	"call_args_mapping": call_args_mapping,	588	"call_args_mapping": call_args_mapping,
496	"mismatch_indices_shapes": mismatch_indices_shapes,	589	"mismatch_indices_shapes": mismatch_indices_shapes,
		590	+ "scalar_resolved_from": scalar_resolved_from,
497	}	591	}
498		592
		593	+
499	def generate_fx_graph_code(code, kernel_code, kernel_name, compile_kwargs):	594	def generate_fx_graph_code(code, kernel_code, kernel_name, compile_kwargs):
500	code = textwrap.indent(code, ' ')	595	code = textwrap.indent(code, ' ')
501	code_template = f"""	596	code_template = f"""
	@@ -546,6 +641,7 @@ num_inputs = {compile_kwargs['num_inputs']}
546	num_outputs = {compile_kwargs['num_outputs']}	641	num_outputs = {compile_kwargs['num_outputs']}
547	non_contiguous_indices = {compile_kwargs['non_contiguous_indices']}	642	non_contiguous_indices = {compile_kwargs['non_contiguous_indices']}
548	mismatch_indices_shapes = {compile_kwargs['mismatch_indices_shapes']}	643	mismatch_indices_shapes = {compile_kwargs['mismatch_indices_shapes']}
		644	+scalar_resolved_from = {compile_kwargs.get('scalar_resolved_from', {})}
549		645
550	async_compile = AsyncCompile()	646	async_compile = AsyncCompile()
551	{kernel_name} = async_compile.triton('{kernel_name}', '''	647	{kernel_name} = async_compile.triton('{kernel_name}', '''
	@@ -567,7 +663,12 @@ def run():
567	args = [arg.npu() if isinstance(arg, torch.Tensor) else arg for arg in args]	663	args = [arg.npu() if isinstance(arg, torch.Tensor) else arg for arg in args]
568		664
569	fx_args = []	665	fx_args = []
570	- for idx in call_args_mapping:	666	+ for pos, idx in enumerate(call_args_mapping):
		667	+ if idx == -1:
		668	+ # SymInt 标量输入: 用编译期记录的精确配方 (kernel_arg_idx, dim) 反推
		669	+ kidx, d = scalar_resolved_from[pos]
		670	+ fx_args.append(int(args[kidx].shape[d]))
		671	+ continue
571	arg = args[idx]	672	arg = args[idx]
572	if isinstance(arg, int):	673	if isinstance(arg, int):
573	fx_args.append(arg)	674	fx_args.append(arg)

Mtorch_npu/_inductor/lowering_patch.py+10-1

	@@ -7,10 +7,10 @@
7	# independently.	7	# independently.
8		8
9	from __future__ import annotations	9	from __future__ import annotations
10	-
11	import copy	10	import copy
12	import functools	11	import functools
13	import importlib	12	import importlib
		13	+import os
14	from collections.abc import Iterable, Mapping	14	from collections.abc import Iterable, Mapping
15	from dataclasses import dataclass, field	15	from dataclasses import dataclass, field
16	from typing import Any, Callable, Optional	16	from typing import Any, Callable, Optional
	@@ -156,6 +156,15 @@ def _uses_device_lowering(
156	)	156	)
157	if layout_device_type is not None:	157	if layout_device_type is not None:
158	return layout_device_type == device_type	158	return layout_device_type == device_type
		159	+ if os.environ.get("INDUCTOR_ASCEND_DUMP_FX_GRAPH") or os.environ.get("INDUCTOR_ASCEND_CHECK_ACCURACY"):
		160	+ device_kwarg = kwargs.get("device")
		161	+ if isinstance(device_kwarg, (torch.device, str)):
		162	+ try:
		163	+ kwarg_device_type = torch.device(device_kwarg).type
		164	+ except Exception:
		165	+ kwarg_device_type = None
		166	+ if kwarg_device_type is not None:
		167	+ return kwarg_device_type in ("npu", "cpu")
159	if device_type in _iter_ir_device_types((args, kwargs)):	168	if device_type in _iter_ir_device_types((args, kwargs)):
160	return True	169	return True
161	return extra_device_predicate is not None and extra_device_predicate(	170	return extra_device_predicate is not None and extra_device_predicate(

Mtorch_npu/_inductor/npu_compare.py+78-3

	@@ -1,12 +1,19 @@
1	import importlib	1	import importlib
		2	+import logging
2	import os	3	import os
3	import sys	4	import sys
		5	+from collections import defaultdict
4	from typing import Any, Iterable, Mapping	6	from typing import Any, Iterable, Mapping
5		7
6	import torch	8	import torch
7	from torch._inductor.compile_fx import clone_preserve_strides	9	from torch._inductor.compile_fx import clone_preserve_strides
8		10
9		11
		12	+log = logging.getLogger(__name__)
		13	+
		14	+_call_counter: dict = defaultdict(int)
		15	+
		16	+
10	def clone_for_accuracy(arg):	17	def clone_for_accuracy(arg):
11	if not isinstance(arg, torch.Tensor):	18	if not isinstance(arg, torch.Tensor):
12	return arg	19	return arg
	@@ -51,14 +58,27 @@ def _report_mismatch(idx, actual, expected, matches, rtol, atol, kernel_name, du
51	rel_diff.masked_fill_(matches, 0)	58	rel_diff.masked_fill_(matches, 0)
52	number_of_elements = matches.numel()	59	number_of_elements = matches.numel()
53	total_mismatches = number_of_elements - int(torch.sum(matches))	60	total_mismatches = number_of_elements - int(torch.sum(matches))
		61	+ mismatch_mask = ~matches
		62	+ mismatch_indices = torch.nonzero(mismatch_mask.flatten())[:5].reshape(-1)
		63	+ sample_info = ""
		64	+ if mismatch_indices.numel() > 0:
		65	+ idx_flat = mismatch_indices[:3].tolist()
		66	+ idx_unflat = [list(torch.unravel_index(torch.tensor(i), actual.shape)) for i in idx_flat]
		67	+ samples = "; ".join(
		68	+ f"pos{idx_unflat[j]}: actual={actual.flatten()[i].item():.6e}, expected={expected.flatten()[i].item():.6e}"
		69	+ for j, i in enumerate(idx_flat)
		70	+ )
		71	+ sample_info = f", Sample values: [{samples}]"
54	msg = (	72	msg = (
55	"CHECK ACCURACY FAILED! "	73	"CHECK ACCURACY FAILED! "
56	f"Kernel: {kernel_name}, Output idx: {idx}, "	74	f"Kernel: {kernel_name}, Output idx: {idx}, "
		75	+ f"actual_shape={list(actual.shape)}, actual_dtype={actual.dtype}, "
		76	+ f"expected_shape={list(expected.shape)}, expected_dtype={expected.dtype}, "
57	f"Mismatched: {total_mismatches}/{number_of_elements} "	77	f"Mismatched: {total_mismatches}/{number_of_elements} "
58	f"({total_mismatches / number_of_elements:.1%}), "	78	f"({total_mismatches / number_of_elements:.1%}), "
59	f"Greatest Rel Diff: {rel_diff.max().item()}, "	79	f"Greatest Rel Diff: {rel_diff.max().item()}, "
60	f"Greatest Abs Diff: {abs_diff.max().item()}, "	80	f"Greatest Abs Diff: {abs_diff.max().item()}, "
61	- f"rtol: {rtol}, atol: {atol}"	81	+ f"rtol: {rtol}, atol: {atol}{sample_info}"
62	)	82	)
63	if dump_path:	83	if dump_path:
64	msg += f", dump_path: {dump_path}"	84	msg += f", dump_path: {dump_path}"
	@@ -117,8 +137,23 @@ def check_accuracy_triton(args, launcher, grid, stream, inductor_meta, *kwargs
117	return None	137	return None
118	call_outputs_indices = fx_module.call_args_mapping[fx_module.num_inputs:]	138	call_outputs_indices = fx_module.call_args_mapping[fx_module.num_inputs:]
119		139
		140	+ _call_counter[kernel_name] += 1
		141	+ invocation = _call_counter[kernel_name]
		142	+ scalar_resolved_from = getattr(fx_module, "scalar_resolved_from", {})
		143	+
		144	+ # 诊断: 记录配方路径的解析值, 供 compare_outputs 失败时打印 (区分误报)
		145	+ _recipe_used = bool(scalar_resolved_from)
		146	+ _recipe_resolved = {}
120	fx_args = []	147	fx_args = []
121	- for idx in fx_module.call_args_mapping:	148	+ for pos, idx in enumerate(fx_module.call_args_mapping):
		149	+ if idx == -1:
		150	+ # SymInt 标量输入: 用编译期记录的精确配方 (kernel_arg_idx, dim) 反推其值。
		151	+ # 配方在 create_compile_kwargs 求解, 序列化为具体 int, 不依赖 pos==dim 假设。
		152	+ kidx, d = scalar_resolved_from[pos]
		153	+ _resolved = int(args[kidx].shape[d])
		154	+ _recipe_resolved[pos] = (scalar_resolved_from[pos], _resolved)
		155	+ fx_args.append(_resolved)
		156	+ continue
122	arg = args[idx]	157	arg = args[idx]
123	if isinstance(arg, int):	158	if isinstance(arg, int):
124	fx_args.append(arg)	159	fx_args.append(arg)
	@@ -130,11 +165,24 @@ def check_accuracy_triton(args, launcher, grid, stream, inductor_meta, *kwargs
130	arg)	165	arg)
131	fx_args.append(fx_arg)	166	fx_args.append(fx_arg)
132		167
		168	+ input_indices = fx_module.call_args_mapping[:fx_module.num_inputs]
		169	+ input_shapes = {f"args[{i}]": list(args[i].shape) if isinstance(args[i], torch.Tensor) else type(args[i]).__name__
		170	+ for i in input_indices}
		171	+ output_indices = call_outputs_indices
		172	+ output_shapes = {f"args[{i}]": list(args[i].shape) if isinstance(args[i], torch.Tensor) else type(args[i]).__name__
		173	+ for i in output_indices}
		174	+ log.debug("[ACC_DEBUG] kernel=%s invocation=#%s inputs=%s outputs=%s",
		175	+ kernel_name, invocation, input_shapes, output_shapes)
		176	+ _input_snapshot = {}
		177	+ for idx in fx_module.call_args_mapping[:fx_module.num_inputs]:
		178	+ if isinstance(args[idx], torch.Tensor):
		179	+ _input_snapshot[idx] = args[idx].cpu()
		180	+
133	fx_graph_call(*fx_args)	181	fx_graph_call(*fx_args)
134		182
135	launcher(args, *kwargs, stream=stream)	183	launcher(args, *kwargs, stream=stream)
136		184
137	- compare_outputs(	185	+ passed = compare_outputs(
138	[args[i] for i in call_outputs_indices],	186	[args[i] for i in call_outputs_indices],
139	fx_args[fx_module.num_inputs:],	187	fx_args[fx_module.num_inputs:],
140	kernel_name=kernel_name,	188	kernel_name=kernel_name,
	@@ -142,6 +190,33 @@ def check_accuracy_triton(args, launcher, grid, stream, inductor_meta, *kwargs
142	dump_path=dump_path,	190	dump_path=dump_path,
143	)	191	)
144		192
		193	+ if not passed and dump_path:
		194	+ fail_path = os.path.join(dump_path, f'data_fail_{invocation}.pth')
		195	+ fail_args = list(args)
		196	+ for idx, cpu_tensor in _input_snapshot.items():
		197	+ fail_args[idx] = cpu_tensor
		198	+ torch.save(fail_args, fail_path)
		199	+ log.warning("[ACC_DEBUG] kernel=%s invocation=#%s FAIL: saved input snapshot to data_fail_%s.pth",
		200	+ kernel_name, invocation, invocation)
		201	+
		202	+ if not passed and _recipe_used:
		203	+ kernel_int_values = [arg for arg in args if isinstance(arg, int)]
		204	+ for pos, ((kidx, d), val) in _recipe_resolved.items():
		205	+ if val in kernel_int_values:
		206	+ log.warning(
		207	+ "[ACC_DEBUG] kernel=%s FAIL: scalar pos=%s resolved=%s "
		208	+ "(from args[%s].shape[%s]), matches kernel int arg. "
		209	+ "Likely a real kernel bug. dump_path=%s",
		210	+ kernel_name, pos, val, kidx, d, dump_path,
		211	+ )
		212	+ else:
		213	+ log.warning(
		214	+ "[ACC_DEBUG] kernel=%s FAIL: scalar pos=%s resolved=%s "
		215	+ "(from args[%s].shape[%s]), NOT found in kernel int args %s. "
		216	+ "Likely a tool artifact (scalar resolution mismatch). dump_path=%s",
		217	+ kernel_name, pos, val, kidx, d, kernel_int_values, dump_path,
		218	+ )
		219	+
145	for arg in fx_args:	220	for arg in fx_args:
146	del arg	221	del arg
147	return True	222	return True

Mtorch_npu/_inductor/runtime/triton_heuristics.py+24-2

	@@ -1790,8 +1790,23 @@ class NPUCachingAutotuner(CachingAutotuner):
1790	if dump_path is None:	1790	if dump_path is None:
1791	log.warning("data dump for kernel %s failed, no valid dump_path is supplied.", self.get_fn_name())	1791	log.warning("data dump for kernel %s failed, no valid dump_path is supplied.", self.get_fn_name())
1792	return False	1792	return False
		1793	+ if not hasattr(self, '_dump_counter'):
		1794	+ self._dump_counter = 0
		1795	+
1793	data_dump_path = os.path.join(dump_path, 'data.pth')	1796	data_dump_path = os.path.join(dump_path, 'data.pth')
1794	- torch.save(args, data_dump_path)	1797	+ input_info = {
		1798	+ i: list(arg.shape) if isinstance(arg, torch.Tensor) else type(arg).__name__
		1799	+ for i, arg in enumerate(args)
		1800	+ }
		1801	+
		1802	+ if self._dump_counter == 0:
		1803	+ log.info("[ACC_DEBUG] %s invocation=#1 saving args=%s to data.pth",
		1804	+ self.get_fn_name(), input_info)
		1805	+ torch.save(args, data_dump_path)
		1806	+ self._dump_counter += 1
		1807	+ else:
		1808	+ log.debug("[ACC_DEBUG] %s invocation=#%s skip save (data.pth preserved from invocation #1, args=%s)",
		1809	+ self.get_fn_name(), self._dump_counter + 1, input_info)
1795	return True	1810	return True
1796		1811
1797	def get_fn_name(self):	1812	def get_fn_name(self):
	@@ -1820,7 +1835,14 @@ class NPUCachingAutotuner(CachingAutotuner):
1820	"Please disable aclgraph before enabling INDUCTOR_ASCEND_CHECK_ACCURACY "	1835	"Please disable aclgraph before enabling INDUCTOR_ASCEND_CHECK_ACCURACY "
1821	"/ INDUCTOR_ASCEND_DUMP_FX_GRAPH, or unset these environment variables."	1836	"/ INDUCTOR_ASCEND_DUMP_FX_GRAPH, or unset these environment variables."
1822	)	1837	)
1823	- _ = self.data_dump(*args)	1838	+ # Strip runtime_block args from the end before saving for replay.
		1839	+ # data_dump receives *launch_args (= kernel_args + runtime_blocks),
		1840	+ # but the replayed script will call run() which re-appends runtime_blocks
		1841	+ # via _build_runtime_launch_args. Saving only kernel_args avoids a
		1842	+ # double-append that overflows positional parameters into "stream".
		1843	+ num_rb = len(self.runtime_block_arg_names or ())
		1844	+ args_for_dump = args[:-num_rb] if num_rb else args
		1845	+ _ = self.data_dump(*args_for_dump)
1824		1846
1825	if npu_config.check_accuracy:	1847	if npu_config.check_accuracy:
1826	if check_accuracy_triton(	1848	if check_accuracy_triton(