Provenance Audit: SimpleMLP — Metadata Across the Compilation Pipeline

Generated: 2026-06-24 19:17  |  Issue: torch-spyre#2574

Measured in-process during one cache-defeated torch.compile (compile-path objects only). This report is measurement-only; interpretation is a separate deliverable.

Quantity	Value
FX pre-grad compute nodes	3
FX post-grad compute nodes	7
LoopLevelIR operations	5
OpSpec ops created	5
SuperDSC kernels	2
sdsc_*.json files	5
OpSpec declared fields	['op', 'is_reduction', 'iteration_space', 'args', 'op_info', 'tiled_symbols']

Stage × Field Matrix

✅ present & non-empty on all instances   ◐ on some (n/total)   ❌ reachable here but measured empty/absent   ➖ not applicable here (no such slot, or carried indirectly via other fields).

Every column tests population (the field exists and carries non-empty content; 0 counts as content, None/[]/{}/"" do not). These cells are measurements only; interpreting each absence is the separate analysis deliverable (provenance_analysis.md).

The Layer column marks whether a field lives on the FX node (FX) or the IR ComputedBuffer (IR). The two IR columns are the same LoopLevelIR before and after the Spyre pre-scheduling passes: LoopLevelIR (pre-pass) is the lowered IR entering them, LoopLevelIR (post-pass) is after they mutate it in place (e.g. inserting restickify buffers). These map to issue #2574’s “Inductor passes” → “LoopLevelIR”.

Layer	Field	FX Graph (pre-grad)	FX Graph (post-grad)	LoopLevelIR (pre-pass)	LoopLevelIR (post-pass)	OpSpec	SuperDSC JSON
FX	stack_trace	✅	◐ 3/7	➖	➖	➖	➖
FX	nn_module_stack	◐ 2/3	◐ 2/7	➖	➖	➖	➖
FX	source_fn_stack	✅	◐ 3/7	➖	➖	➖	➖
FX	original_aten	➖	✅	➖	➖	➖	➖
FX	from_node	➖	◐ 3/7	➖	➖	➖	➖
IR	origins	➖	➖	✅	✅	❌	❌
IR	origin_node	➖	➖	✅	✅	❌	❌
IR	traceback	➖	➖	❌	❌	❌	❌
IR	get_stack_traces	➖	➖	◐ 3/5	◐ 3/5	❌	❌

Stage 2 — FX Graph (pre-grad): 3 compute nodes

Cell = observed type of the field, or ❌ if absent.

Node	target	stack_trace	nn_module_stack	source_fn_stack	original_aten	from_node	source line
x	<built-in function linear>	str	dict	list	❌	❌	x = self.fc1(x)
x_1	<built-in method relu of type object at 0x7f9edd8058a0>	str	❌	list	❌	❌	x = torch.relu(x)
x_2	<built-in function linear>	str	dict	list	❌	❌	x = self.fc2(x)

Stage 2 — FX Graph (post-grad): 7 compute nodes

Cell = observed type of the field, or ❌ if absent.

Node	target	stack_trace	nn_module_stack	source_fn_stack	original_aten	from_node	source line
permute	aten.permute.default	str	dict	list	OpOverload	list	x = self.fc1(x)
mm_default_1	aten.mm.default	❌	❌	❌	OpOverload	❌	—
add_tensor_1	aten.add.Tensor	❌	❌	❌	OpOverload	❌	—
relu	aten.relu.default	str	❌	list	OpOverload	list	x = torch.relu(x)
permute_1	aten.permute.default	str	dict	list	OpOverload	list	x = self.fc2(x)
mm_default	aten.mm.default	❌	❌	❌	OpOverload	❌	—
add_tensor	aten.add.Tensor	❌	❌	❌	OpOverload	❌	—

Stage 3 — LoopLevelIR (pre-pass): 5 operations

The lowered IR entering the Spyre pre-scheduling passes.

Op	origins	origin_node	traceback	get_stack_traces
op0	mm_default_1, permute	mm_default_1	❌	✅
op1	add_tensor_1	add_tensor_1	❌	❌
op2	relu	relu	❌	✅
op3	mm_default, permute_1	mm_default	❌	✅
op4	add_tensor	add_tensor	❌	❌

Stage 4 — LoopLevelIR (post-pass): 5 operations

The same IR after the pre-scheduling passes mutate it in place.

Op	origins	origin_node	traceback	get_stack_traces
op0	mm_default_1, permute	mm_default_1	❌	✅
op1	add_tensor_1	add_tensor_1	❌	❌
op2	relu	relu	❌	✅
op3	mm_default, permute_1	mm_default	❌	✅
op4	add_tensor	add_tensor	❌	❌

Stage 5 — OpSpec: 5 ops

OpSpec declared fields: ['op', 'is_reduction', 'iteration_space', 'args', 'op_info', 'tiled_symbols'] — no provenance field. The origins below are what is available on the input ComputedBuffer at create_op_spec; the OpSpec object itself declares no field to hold them.

Spyre op	buffer	origins	origin_node
batchmatmul	op0	mm_default_1, permute	mm_default_1
add	op1	add_tensor_1	add_tensor_1
relufwd	op2	relu	relu
batchmatmul	op3	mm_default, permute_1	mm_default
add	op4	add_tensor	add_tensor

Stage 6 — SuperDSC: 5 sdsc_*.json files (2 kernels)

Provenance field present in any emitted sdsc_*.json: ❌

sdsc_fused_addmm_linear_relu_0

• buffers (4): op0, op1, op2, op3

• fx origins: add_tensor_1, mm_default, mm_default_1, permute, permute_1, relu

• kernel metadata: # Topologically Sorted Source Nodes: [x, x_1, x_2], Original ATen: [aten.linear, aten.addmm, aten.relu]

• sdsc_*.json files: 4   provenance in JSON: ❌

sdsc_fused_addmm_1

• buffers (1): op4

• fx origins: add_tensor

• kernel metadata: # Topologically Sorted Source Nodes: [], Original ATen: [aten.addmm]

• sdsc_*.json files: 1   provenance in JSON: ❌