DLA surgeries added as part of ModelOpt#1415
Conversation
Signed-off-by: mgohil-png <mgohil@nvidia.com>
📝 WalkthroughWalkthroughAdds shared ONNX utilities, numerous DLA-focused graph transforms, a centralized surgery runner, a multi-stage DLA compatibility pipeline, refactors to in-memory transforms with wrappers, and extensive pytest suites validating structural rewrites and ORT parity. ChangesDLA Graph Surgery and Orchestration
Sequence Diagram(s)sequenceDiagram
autonumber
participant CLI as User/CLI
participant RG as run_graph_surgery
participant REG as SurgeryRegistry
participant TR as TransformFn
participant IO as Loader/Saver
CLI->>RG: request(surgery_name, model_path, output_path, kwargs)
RG->>IO: load ONNX model (+external data)
RG->>REG: resolve surgery_name→transform
RG->>TR: call(model, verbose, kwargs)
TR-->>RG: modified model
RG->>IO: save model (+external data, infer shapes)
IO-->>CLI: output_path
Estimated code review effort🎯 5 (Critical) | ⏱️ ~120 minutes ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
|
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Note
Due to the large number of review comments, Critical severity comments were prioritized as inline comments.
🟠 Major comments (27)
modelopt/onnx/graph_surgery/dla_transforms/dla_remove_qdq.py-194-199 (1)
194-199:⚠️ Potential issue | 🟠 Major | ⚡ Quick winPotential bug: only the first matching input is updated for multi-input consumers.
If a consumer node uses
qdq_node_pair_inputin multiple input slots (e.g.,Add(x, x)orConcatwith repeated inputs), thebreakstatement causes only the first occurrence to be renamed tograph_output.name, leaving subsequent references dangling.🐛 Proposed fix: remove the break to update all matching inputs
for j, input_name in enumerate(consumer.input): if input_name == qdq_node_pair_input: consumer.input[j] = graph_output.name - break🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_remove_qdq.py` around lines 194 - 199, The loop that updates consumer inputs only replaces the first matching occurrence of qdq_node_pair_input because of the break inside the inner loop; remove the break so all input slots that equal qdq_node_pair_input are updated to graph_output.name. Locate the block using consumer_node_map.get(qdq_node_pair_input, []), iterate over each consumer and its consumer.input list, and replace every matching input_name with graph_output.name (no early exit) to handle multi-input consumers like Add(x,x) or repeated Concat entries.modelopt/onnx/graph_surgery/dla_transforms/dla_constants_to_initializers.py-50-52 (1)
50-52:⚠️ Potential issue | 🟠 Major | ⚡ Quick win
sparse_valuepath is using the wrong ONNX proto type and will crash conversion.At line 51,
numpy_helper.to_array()is called withattr.sparse_tensor(SparseTensorProto), but that API is documented to accept onlyTensorProto. Passing aSparseTensorProtowill raiseAttributeErrororTypeError, which bypass yourexcept ValueErrorguard (line 77), aborting the transform for any model usingsparse_valueconstants.🐛 Proposed fix
if attr.name == "sparse_value": - arr = numpy_helper.to_array(attr.sparse_tensor) - return numpy_helper.from_array(arr, name=out_name) + sparse = attr.sparse_tensor + values = numpy_helper.to_array(sparse.values) + indices = numpy_helper.to_array(sparse.indices).astype(np.int64) + dense = np.zeros(tuple(sparse.dims), dtype=values.dtype) + if indices.ndim == 1: + dense.reshape(-1)[indices] = values + else: + dense[tuple(indices.T)] = values + return numpy_helper.from_array(dense, name=out_name)- except ValueError as exc: + except (ValueError, TypeError, AttributeError) as exc: if verbose: logger.warning("%s", exc) continueAlso applies to: 77-80
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_constants_to_initializers.py` around lines 50 - 52, The sparse_value branch is incorrectly calling numpy_helper.to_array on attr.sparse_tensor (a SparseTensorProto); instead extract the SparseTensorProto fields, convert the sparse values and indices to numpy arrays using numpy_helper.to_array(attr.sparse_tensor.values) and numpy_helper.to_array(attr.sparse_tensor.indices), allocate a dense numpy array of shape attr.sparse_tensor.dims with the proper dtype, scatter the values into the dense array at the indices (e.g. dense[tuple(indices.T)] = values), and then return numpy_helper.from_array(dense, name=out_name); apply the same pattern to the similar code path around the later block (the code around the except/lines 77-80) that also handles sparse_value.modelopt/onnx/graph_surgery/dla_transforms/dla_unsqueeze.py-28-49 (1)
28-49:⚠️ Potential issue | 🟠 Major | ⚡ Quick winPotential
IndexErrorwhen Unsqueeze has no consumers.At line 46,
node_input_pair[0][0]will raise anIndexErrorifget_consumers(node.output[0])returns an empty list. This can happen if the Unsqueeze output is a graph output or is otherwise unused.Suggested fix
def _apply_unsqueeze(model): """Transform Unsqueeze node when data is of type int32 or int64.""" cache = GraphCache(model.graph) nodes_to_add_at_pos = [] for node in cache.nodes_by_op("Unsqueeze"): data_dtype = cache.get_dtype(node.input[0]) if data_dtype in (TensorProto.INT32, TensorProto.INT64): + node_input_pair = cache.get_consumers(node.output[0]) + if not node_input_pair: + # No consumers to rewire; skip this node + continue cast_output_name = f"{node.output[0]}_float32" cast_node = helper.make_node( "Cast", inputs=[node.output[0]], outputs=[cast_output_name], to=TensorProto.FLOAT, name=f"{node.output[0]}_cast_to_float32", ) - node_input_pair = cache.get_consumers(node.output[0]) for consumer, idx in node_input_pair: consumer.input[idx] = cast_output_name nodes_to_add_at_pos.append(([cast_node], node_input_pair[0][0])) for nodes_to_add, node_pos in nodes_to_add_at_pos: insert_nodes_at_position(model.graph, nodes_to_add, node_pos) return model🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_unsqueeze.py` around lines 28 - 49, The Unsqueeze transform in _apply_unsqueeze assumes cache.get_consumers(node.output[0]) returns at least one consumer, causing an IndexError when the output has no consumers (e.g., graph output); change the logic to handle an empty node_input_pair: after computing node_input_pair = cache.get_consumers(...), only iterate and rewrite consumer.input when node_input_pair is non-empty, and choose node_pos = node_input_pair[0][0] when present but fall back to the Unsqueeze node itself (the original node variable) when node_input_pair is empty before appending to nodes_to_add_at_pos and before calling insert_nodes_at_position; this avoids indexing into an empty list and ensures insertion still occurs in the correct place.modelopt/onnx/graph_surgery/dla_transforms/_dla_graph_helpers.py-178-198 (1)
178-198:⚠️ Potential issue | 🟠 Major | ⚡ Quick winPotential crash when
x_zero_pointisNone.At lines 185-190,
x_zero_point.dtypeis accessed to determineint_maxandint_minbefore theNonecheck at line 191. Ifget_initializer_by_namereturnsNone, this will raise anAttributeError.Suggested fix
def calculate_clip_range(node, model): x_scale = get_initializer_by_name(model, node.input[1]) if x_scale is None: x_scale = get_constant_by_name(model, node.input[1]) x_zero_point = get_initializer_by_name(model, node.input[2]) if x_scale is None: raise ValueError(f"{node.name} should have x_scale value") + if x_zero_point is None: + logger.info("x_zero_point is None!") + x_zero_point = np.array(0, dtype=np.int32) + int_max = np.int32(127) + int_min = np.int32(-128) + else: - int_max = np.int32( - 65535 if x_zero_point.dtype == np.uint16 else 255 if x_zero_point.dtype == np.uint8 else 127 - ) - int_min = np.int32( - 0 if x_zero_point.dtype == np.uint16 else 0 if x_zero_point.dtype == np.uint8 else -128 - ) - if x_zero_point is None: - logger.info("x_zero_point is None!") - x_zero_point = np.array(0, dtype=np.int32) - else: + int_max = np.int32( + 65535 if x_zero_point.dtype == np.uint16 else 255 if x_zero_point.dtype == np.uint8 else 127 + ) + int_min = np.int32( + 0 if x_zero_point.dtype == np.uint16 else 0 if x_zero_point.dtype == np.uint8 else -128 + ) x_zero_point = x_zero_point.astype(np.int32) clip_min = ((int_min - x_zero_point) * x_scale).astype(np.float32) clip_max = ((int_max - x_zero_point) * x_scale).astype(np.float32) return clip_min, clip_max🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/_dla_graph_helpers.py` around lines 178 - 198, calculate_clip_range accesses x_zero_point.dtype before checking for None which can crash; modify the function so the None check for x_zero_point happens immediately after retrieving it (and before computing int_max/int_min), e.g. if x_zero_point is None set it to np.array(0, dtype=np.int32) (or otherwise ensure a valid dtype), then cast x_zero_point to np.int32 and only then compute int_max and int_min and the final clip_min/clip_max; keep references to the same symbols: calculate_clip_range, x_scale, x_zero_point, int_max, int_min.modelopt/onnx/graph_surgery/dla_transforms/dla_not.py-76-110 (1)
76-110:⚠️ Potential issue | 🟠 Major | ⚡ Quick winUse a fallback id when
node.nameis empty.
node.nameis optional in ONNX. For anonymousNotnodes, this generates the same*_clip_min,*_clip_max,*_one, and*_clippedtensor names on every rewrite, so multiple matches collide and can invalidate the graph.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_not.py` around lines 76 - 110, The code currently builds names using node.name which can be empty; change the naming to use a fallback unique id when node.name is falsy (e.g., use node.output[0] or a generated UUID/counter) before constructing min_name, max_name, one_name, clip_out and the node.name used for clip_node/sub_node; ensure add_unique_initializers, helper.make_node calls and the name used to create value_info all use that fallback identifier so anonymous Not nodes produce unique tensor and node names on each rewrite.modelopt/onnx/graph_surgery/dla_transforms/dla_remove_reshapes.py-39-44 (1)
39-44:⚠️ Potential issue | 🟠 Major | ⚡ Quick winSkip reshape-chain folding when endpoint shapes aren't concrete.
get_tensor_shape_by_name(...)can be unknown or symbolic on dynamic models. Converting that straight intonp.array(..., dtype=np.int64)will raise before any rewrite happens. Guard both shape lookups and leave the chain untouched when either endpoint shape is not fully resolved.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_remove_reshapes.py` around lines 39 - 44, The code converts endpoint shapes from get_tensor_shape_by_name(...) directly into np.array(dtype=np.int64), which will raise on unknown/symbolic shapes; update the logic where reshape_chain endpoints are fetched (use reshape_chain[0].input[0] and reshape_chain[-1].output[0] via get_tensor_shape_by_name(model.graph, ...)) to first verify each returned shape is concrete (not None and every dimension is an integer/fully resolved) and only then build np.array(dtype=np.int64); if either endpoint shape is non‑concrete, skip folding the reshape_chain and leave it untouched.modelopt/onnx/graph_surgery/dla_transforms/dla_squeeze_unsqueeze_to_reshape.py-67-89 (1)
67-89:⚠️ Potential issue | 🟠 Major | ⚡ Quick winValidate the axes initializer before reading it, and require an exact one-axis pattern.
numpy_helper.to_array(...)is called before theNonecheck, so non-initializer axes inputs fail with an opaque exception. Also, onlyaxes_array[0]is validated, which lets multi-axis cases like[0, 2]slip through and be rewritten incorrectly.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_squeeze_unsqueeze_to_reshape.py` around lines 67 - 89, In transform_squeeze_unsqueeze_to_reshape, validate initializer objects before calling numpy_helper.to_array and enforce a single-axis pattern: first ensure squeeze_axes = get_initializer_object_by_name(model, node.input[1]) is not None before converting, then convert to array and require squeeze_axes_array.ndim == 1 and squeeze_axes_array.size == 1 and squeeze_axes_array[0] in [0,1,2]; do the same for unsqueeze_axes (ensure consumer.input[1] initializer exists, then to_array, then require size == 1 and unsqueeze_axes_array[0] == -1). Use the existing variable names (squeeze_axes, squeeze_axes_array, unsqueeze_axes, unsqueeze_axes_array) and functions (get_initializer_object_by_name, numpy_helper.to_array) to locate and apply the checks.modelopt/onnx/graph_surgery/dla_transforms/dla_greater.py-79-89 (1)
79-89:⚠️ Potential issue | 🟠 Major | ⚡ Quick winDeduplicate inserted Casts for shared tensors.
cast_out = f"{inp}_to_float32"is reused blindly. If the same non-float tensor feeds both inputs of oneGreateror multipleGreaternodes, this will create duplicateCastnodes/output names. Reuse an existing cast or uniquify the generated names.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_greater.py` around lines 79 - 89, The code blindly creates a Cast node named via cast_out/ cast_node for any non-float input, causing duplicate Casts when the same tensor is reused; before creating a new Cast for inp in the Greater transform, search the existing graph nodes for an output matching cast_out (or maintain a local map of already-created casts keyed by inp) and if found reuse that output name instead of creating a new node; only call insert_nodes_at_position(graph, [cast_node], node) and assign node.input[idx] = cast_out when no existing cast was found, otherwise set node.input[idx] to the existing cast's output; reference cast_out, cast_node, insert_nodes_at_position, and node.input in the Greater transform.modelopt/onnx/graph_surgery/dla_transforms/dla_not.py-60-74 (1)
60-74:⚠️ Potential issue | 🟠 Major | ⚡ Quick winAvoid rewriting a shared Cast in place.
This changes the producer
CasttoFLOATfor every matchingNot. If that Cast output fans out to another bool consumer, the other branch now receivesFLOATinstead ofBOOL. Either require a single consumer here or clone the Cast for the rewritten path.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_not.py` around lines 60 - 74, The current loop over cache.nodes_by_op("Not") mutates the producer Cast node in place (producer_node) by changing its "to" attribute and its output value_info (cast_out_vi), which breaks other consumers; instead, for each Not node ensure the Cast has a single consumer or create a cloned Cast node for this Not path: if cache.get_consumers(producer_node.output[0]) returns exactly one (the Not), you may safely change the existing producer_node.attribute 'to' and its cast_out_vi.elem_type to TensorProto.FLOAT; otherwise, create a new Cast node (clone of producer_node with 'to' set to FLOAT), insert it into the graph, update the Not node.input to point to the clone.output, and add/update the cloned output value_info via get_tensor_value_info_by_name(model, clone.output[0]) to TensorProto.FLOAT so the original Cast and its other consumers remain unchanged.modelopt/onnx/graph_surgery/dla_transforms/dla_remove_deqlin.py-81-97 (1)
81-97:⚠️ Potential issue | 🟠 Major | ⚡ Quick winDon't bypass the protected-weight skip rules on fan-out.
When
dq_outputhas more than one consumer,_should_skip()returnsFalseimmediately. That means aDequantizeLinearfeeding a protected weight use likeConv/MatMulplus any second consumer will still be folded, even though the logic below explicitly says that weight path must stay quantized. Check all consumers for protected uses before deciding to remove the DQ.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_remove_deqlin.py` around lines 81 - 97, The current logic returns False immediately when dq_output has multiple consumers, which lets a DequantizeLinear feeding a protected weight use slip through; instead iterate over all entries in consumer_map.get(dq_output, []) and for each consumer_node/input_idx apply the same protected-weight checks (is_weight = input_idx == 1, op in {"Conv","ConvTranspose","MatMul","Gemm"} and zp_dtype membership in _CONV_SKIP_ZP/_CONVT_SKIP_ZP/_MATMUL_SKIP_ZP). For MatMul/Gemm call check_to_apply_transpose(consumer_node, model, shape_map) inside a try/except and treat any consumer that matches as a reason to skip (return True). Only if none of the consumers are protected should the function continue returning False or proceed with folding.modelopt/onnx/graph_surgery/dla_transforms/dla_handle_qdq.py-101-107 (1)
101-107:⚠️ Potential issue | 🟠 Major | ⚡ Quick winAdjust the implicit Q/DQ axis too.
When
axisis omitted, Q/DQ still uses the default channel axis of1. After prepending singleton dims, that referenced dimension moves, but this code only shifts explicit attributes. Per-channel quantization that relies on the default axis will therefore point at the wrong dimension after wrapping.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_handle_qdq.py` around lines 101 - 107, The code only shifts explicit "axis" attributes but ignores the implicit default axis=1 used by Q/DQ; update the logic in the node attribute handling (the has_axis/orig_axis/set_node_attr_i block) to account for the omitted-attribute case by treating missing axis as 1 and writing back an adjusted axis (use get_node_attr_i(node, "axis", 1) to read and set_node_attr_i(node, "axis", orig_axis + delta) when orig_axis >= 0), i.e., if has_axis is false still compute orig_axis = 1, add delta, and set the attribute so per-channel quantization now references the moved dimension.modelopt/onnx/graph_surgery/dla_transforms/dla_greater.py-63-68 (1)
63-68:⚠️ Potential issue | 🟠 Major | ⚡ Quick winConvert any non-
FLOATinitializer operand, not just INT32/INT64.The docstring says every
Greaterinput should end upFLOAT, but this branch skipsFLOAT16,DOUBLE,INT8, etc. Those static operands remain non-float32 while dynamic tensors are cast, so the transform is incomplete for initializer-backed thresholds.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_greater.py` around lines 63 - 68, The transform currently only casts initializers of integer types because of the check "if dtype not in _INT_TYPES: continue"; change this to treat any non-float32 initializer as eligible for casting: replace the dtype guard with "if dtype == np.float32: continue" (or equivalent) so FLOAT16/DOUBLE/INT8/etc. are not skipped, then call cache.get_init_array(inp), astype(np.float32) and write the new array back into the initializer storage (use the cache's setter method such as cache.set_init_array or the repository's initializer-replace helper) so the initializer backing 'Greater' inputs is actually replaced with a float32 tensor; keep the existing assert init_arr is not None and update any variable names (init_arr/arr) accordingly.modelopt/onnx/graph_surgery/dla_transforms/dla_fix_instancenorm_channel_mismatch.py-112-139 (1)
112-139:⚠️ Potential issue | 🟠 Major | ⚡ Quick winPreserve the original tensor dtype in the injected
value_info.
InstanceNormalizationoutputs the same element type as its input, butinstrnorm_outandpost_outare always recorded asFLOAT. On FP16 graphs that makes the rewritten metadata lie about the intermediate dtype and can confuse later passes that readvalue_info.Suggested fix
- add_value_info(graph, instrnorm_out, onnx.TensorProto.FLOAT, [n, c, d, 1]) + add_value_info(graph, instrnorm_out, in_dtype, [n, c, d, 1]) ... - add_value_info(graph, post_out, onnx.TensorProto.FLOAT, [1, n, c, d]) + add_value_info(graph, post_out, in_dtype, [1, n, c, d])🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_fix_instancenorm_channel_mismatch.py` around lines 112 - 139, The injected value_info for instrnorm_out and post_out always uses onnx.TensorProto.FLOAT, which misreports dtype on FP16 graphs; change both add_value_info calls to use the original input element type (the in_dtype computed via cache.get_dtype(in_name) or equivalent) so InstanceNormalization's output and the subsequent reshape preserve the true dtype metadata (update the add_value_info calls for instrnorm_out and post_out to pass in_dtype rather than onnx.TensorProto.FLOAT).modelopt/onnx/graph_surgery/dla_transforms/dla_cast_to_fp32.py-77-81 (1)
77-81:⚠️ Potential issue | 🟠 Major | ⚡ Quick winUpdate
graph.outputtypes when a rewritten Cast is a model output.This only patches
graph.value_info. Ifnode.output[0]is also declared ingraph.output, the model metadata still reports the old elem type even though the Cast now returnsFLOAT.Suggested fix
- for vi in graph.value_info: + for vi in list(graph.value_info) + list(graph.output): if vi.name == node.output[0]: vi.type.tensor_type.elem_type = TensorProto.FLOAT - break🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_cast_to_fp32.py` around lines 77 - 81, The code updates graph.value_info but misses updating graph.output when the Cast output is also a declared model output; modify the loop (or add a second loop) to check graph.output entries and set their type to TensorProto.FLOAT when vi.name == node.output[0] (i.e., update graph.output[i].type.tensor_type.elem_type as well as graph.value_info) so the model metadata reflects the rewritten Cast output; ensure you reference node.output[0], graph.value_info, and graph.output and set elem_type = TensorProto.FLOAT consistent with the existing change.modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py-247-272 (1)
247-272:⚠️ Potential issue | 🟠 Major | ⚡ Quick winHandle
Transposenodes that rely on the default permutation.ONNX treats a missing
permas reversing the axes. This handler returnsNonewhen the attribute is absent, so default-formTransposenodes stay non-4D and escape the pass entirely.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py` around lines 247 - 272, The code returns None when a Transpose node lacks a "perm" attribute, but ONNX semantics treat a missing perm as the reversed axes; update the handler so if no attr with name "perm" is found you construct perm = list(range(rank-1, -1, -1)) (i.e., default reverse permutation) and then apply the existing rank-specific logic (using rank, orig_shape, idx_to_remove, etc.), set perm_updated = True and write the computed new_perm back into attr.ints (or create the attribute entry before writing) so default-form Transpose nodes are handled like explicit perms.modelopt/onnx/graph_surgery/dla_transforms/dla_decompose_lstm.py-773-915 (1)
773-915:⚠️ Potential issue | 🟠 Major | ⚡ Quick winFail fast on unsupported LSTM features instead of rewriting them with default semantics.
This pass hardcodes the default ONNX LSTM behavior, but it never validates
layout,sequence_lens,P,clip,input_forget, or customactivations/ activation params. Models using any of those features are currently decomposed as if they were absent, which silently changes results.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_decompose_lstm.py` around lines 773 - 915, The transform silently ignores unsupported LSTM features; add explicit validation early in the per-node handling (before transpose_inputs/split_direction_wise) to fail-fast when unsupported attributes are present: inspect node.attribute and check for non-default/unsupported values of "layout", "sequence_lens", "P" (presence), "clip" (non-zero), "input_forget" (true), and custom "activations"/"activation_alpha"/"activation_beta" (anything other than the default activation set); if any unsupported feature is found, raise a clear ValueError referencing the node.name; implement this as a small helper (e.g., validate_lstm_supported(node)) and call it right after computing direction/shape_info_map so the rest of the code (transpose_inputs, split_direction_wise, add_nodes_for_time_step) can assume only the supported default semantics.modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py-76-151 (1)
76-151:⚠️ Potential issue | 🟠 Major | 🏗️ Heavy liftPreserve shared weights instead of reshaping them in place.
Both the eligible path and the fallback path rewrite the original initializer proto directly. If that weight is shared by another
MatMul/Gemmor any unrelated op, the other consumer will now see Conv-shaped data or lose the expected tensor layout entirely. Please clone/repoint single-LSTM consumers instead of mutating shared initializers in place.Also applies to: 187-197
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py` around lines 76 - 151, The update_initializers function currently mutates the original initializer proto in-place which breaks other consumers; instead detect if the initializer (initializer.name / name) has multiple consumers in graph.node and if so clone it: create a new unique tensor name, create a new initializer proto from reshaped_arr via numpy_helper.from_array (do not call initializer.CopyFrom on the original), add the new initializer to graph.initializer, and repoint only the MatMul/Gemm/DequantizeLinear node inputs (the node found by dq_node or the node being transformed) to the new name; when the initializer has a single consumer you may keep the in-place replacement as before. Apply the same cloning logic to the other similar block mentioned (lines ~187-197) so shared weights are never mutated in-place.modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py-989-997 (1)
989-997:⚠️ Potential issue | 🟠 Major | ⚡ Quick winPad
Tile.repeatsfor every promoted rank, not just 3D.This handler unsqueezes any non-4D input to 4D, but it only expands the
repeatsinitializer when its length is 3. A 2DTilewill therefore get a 4D input with a 2-element repeats vector, which is invalid.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py` around lines 989 - 997, The repeats initializer padding only handles len==3; change the logic in the Tile-promoting block (the code that checks node.input[1] in init_names and iterates graph.initializer) to pad any repeats array with leading ones until its length equals 4 instead of only handling 3D: read init via numpy_helper.to_array(init) into arr, if arr.ndim==1 and len(arr) < 4 create new_arr = np.concatenate([np.ones(4-len(arr), dtype=arr.dtype), arr]).astype(arr.dtype) and then init.CopyFrom(numpy_helper.from_array(new_arr, init.name)); keep existing break and other checks intact so 1D/2D/3D repeats are all promoted correctly when promoting inputs to 4D.modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py-185-196 (1)
185-196:⚠️ Potential issue | 🟠 Major | ⚡ Quick winShift the DQ channel axis in the fallback 4D-promotion path.
This branch pads the quantized source initializer to 4D for a
Const -> DequantizeLinearweight, but it never updates the matchingDequantizeLinear.axis. Per-channel scales/zero-points will target the wrong dimension after promotion.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py` around lines 185 - 196, The fallback 4D-promotion pads a const initializer but doesn't adjust any DequantizeLinear.axis, so per-channel scales/zero-points will point to the wrong dim; after you compute arr and new_shape in the branch that reshapes the initializer, find all DequantizeLinear nodes in graph whose input (the quantized input) equals init_name and if they have an 'axis' attribute increment that axis by (len(new_shape) - arr.ndim) (i.e. the number of leading dims you added) — update the attribute in-place on those nodes so the axis refers to the same logical channel after promotion.modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py-248-345 (1)
248-345:⚠️ Potential issue | 🟠 Major | ⚡ Quick winReject non-default
Gemm.alpha/betaor fold them into the rewrite.The conversion only honors
transA,transB, and optional bias.Gemmalso appliesalpha * A'B' + beta * C; any node withalpha != 1orbeta != 1will change numerics after being replaced withConv.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.py` around lines 248 - 345, The transform currently ignores Gemm attributes alpha and beta; inspect node.attribute for "alpha" and "beta" near the top of this rewrite (alongside transA/transB) and if either is not 1.0 either (a) reject the rewrite by raising a ValueError referencing node.name, or (b) fold them into the Conv inputs: multiply the weight initializer by alpha (use update_initializers or modify the initializer named by init_name when is_const_dq is True; for dynamic weights insert a Mul before Conv using conv_input1), and apply beta to the bias/C input (if C is present and constant scale the bias tensor, otherwise insert a Mul on the bias/C input before adding to conv). Ensure the chosen solution handles both static and dynamic weight/bias paths and updates conv_inputs and initializer updates consistently.modelopt/onnx/graph_surgery/dla_transforms/dla_decompose_lstm.py-81-105 (1)
81-105:⚠️ Potential issue | 🟠 Major | 🏗️ Heavy liftDon't delete or rewrite initializers that may still be shared.
These helpers replace/remove the original initializer unconditionally after creating the LSTM-specific 4D/transposed version. If another node still references the original tensor, that consumer is left with a missing name or with a layout that was only valid for this decomposition.
Also applies to: 107-120, 304-349
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_decompose_lstm.py` around lines 81 - 105, The code creates a new 4D LSTM-specific initializer (via _expand_array_to_rank4 and numpy_helper.from_array) and then unconditionally removes the original initializer (graph.initializer.remove(initializer)) which may still be referenced by other nodes; instead, after creating new_init and adding it with add_unique_initializers, check whether any other graph.node (excluding the consumer you just rewired, e.g., deq_node or the LSTM consumer in the other ranges) still references initializer.name — if no other references exist, remove it; otherwise leave the original initializer in place and only set the specific consumer's input to new_init.name. Apply the same pattern for the other affected code blocks (the ranges around 107-120 and 304-349) and use functions/variables like cache.get_init, _expand_array_to_rank4, add_unique_initializers, graph.initializer.remove, and the consumer node inputs to locate the spots to change.modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py-439-475 (1)
439-475:⚠️ Potential issue | 🟠 Major | ⚡ Quick winShift implicit Slice axes after promoting the input to 4D.
When
axesis omitted, ONNX interprets it as[0..len(starts)-1]. After the input is unsqueezed, those implied axes need the same delta as explicit axes; otherwise a 2D/3DSlicewill slice the prepended singleton dims instead of the original data dims.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py` around lines 439 - 475, The Slice transform fails to shift implicit axes (when node.input[3] is missing/empty) after unsqueezing; update the block handling axes so if node.input[3] is absent or not in init_names you read the starts initializer (node.input[1]) to get its length, build an axes array as np.arange(len(starts)), apply delta = _axis_delta(rank) to each axis, create a new initializer with numpy_helper.from_array (reusing cnt_ref and inits_to_add like the explicit-axes path), set node.input[3] to the new init name, and otherwise keep the existing explicit-axes code path (use same symbols: node, node.input, graph.initializer, starts initializer, _axis_delta, cnt_ref, inits_to_add, numpy_helper.from_array).modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py-552-615 (1)
552-615:⚠️ Potential issue | 🟠 Major | ⚡ Quick winTreat axis-less Reduce as “reduce all dims” in the post-squeeze.
If a
Reduce*node omitsaxes,adjusted_axesstays empty here. Forkeepdims=0, the fallback squeeze then removes only the promoted leading dims, so reduce-all cases come back as[1, ...]shapes instead of the original scalar / lower-rank output.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.py` around lines 552 - 615, The post-squeeze currently misses the case where a Reduce node has no axes (meaning "reduce all dims"), so when orig_keepdims==0 and adjusted_axes is empty the code only squeezes promoted leading dims; to fix, when axes are omitted set adjusted_axes = [i + delta for i in range(rank)] (use the existing delta computed by delta = _axis_delta(rank)) before the Squeeze logic so the squeeze will remove all original reduced dimensions; update the block that computes adjusted_axes (the branch after checking node.input[1] and the attribute branch) to populate adjusted_axes accordingly when no axes were found so the later Squeeze/Sq_axes computation (variables adjusted_axes, orig_keepdims, leading, sq_axes, temp_out) behaves correctly.modelopt/onnx/graph_surgery/make_dla_compatible.py-178-220 (1)
178-220:⚠️ Potential issue | 🟠 Major | ⚡ Quick win
dla_make_dla_compatibleno longer matches the exported surgery API.This helper is re-exported as
make_dla_compatible, but unlike the other public surgeries it neither acceptsoutput_pathnor saves the result. The example inmodelopt/onnx/graph_surgery/__init__.pycalls it withoutput_path=..., which will currently raiseTypeError. Please either make this a normal file-to-file wrapper or keep it private and steer callers torun_graph_surgery.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/make_dla_compatible.py` around lines 178 - 220, The public helper dla_make_dla_compatible currently only takes model_path and returns a ModelProto, which mismatches the exported surgery API that expects an output_path/file-to-file wrapper; update dla_make_dla_compatible to accept an optional output_path: Optional[str] = None (and keep verbose) and after calling _transform_make_dla_compatible(model, ...) save the transformed model to output_path when provided (use onnx.save / appropriate external data flags) and still return the ModelProto; reference the existing symbols model, _transform_make_dla_compatible, and logger to locate where to load, transform, optionally save, and return the model.modelopt/onnx/graph_surgery/dla_transforms/dla_5d_reshape_to_4d.py-57-66 (1)
57-66:⚠️ Potential issue | 🟠 Major | ⚡ Quick winReject fan-out before rewriting the follower chain.
_parse_middleware_strict()currently takescands[0]whencurhas multiple first-input consumers. In a branched graph that means this pass rewrites/removes only one branch and leaves the others pointing at deleted tensors or nodes. Please require a single consumer here (or skip the candidate) instead of choosing one arbitrarily.💡 Proposed fix
cands = _consumers_first_input(graph, cur) if not cands: msg = ( f"{path_context}: no consumer for tensor {cur!r}; " f"expected middleware (Clip / Q→DQ) or terminal in {sorted(stop_op_types)}" ) raise ValueError(msg) + if len(cands) != 1: + msg = ( + f"{path_context}: expected a single first-input consumer for tensor {cur!r}, " + f"got {[n.name for n in cands]}" + ) + raise ValueError(msg) n = cands[0]🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_5d_reshape_to_4d.py` around lines 57 - 66, The _parse_middleware_strict() loop currently picks cands[0] from _consumers_first_input(graph, cur) which allows fan-out to be silently ignored and later leaves other branches pointing at removed tensors; change the logic to require exactly one first-input consumer (len(cands) == 1) before treating it as the next follower, otherwise skip this candidate (or treat it as non-middleware) and continue/raise as appropriate so you do not arbitrarily pick cands[0]; update references to cands, cur and the stop_op_types check to only act when a single consumer exists.modelopt/onnx/graph_surgery/make_dla_compatible.py-162-171 (1)
162-171:⚠️ Potential issue | 🟠 Major | ⚡ Quick winDon't continue after
onnx.checker.ValidationError.If the transformed graph fails ONNX validation, returning it as “done” makes the failure surface later during save/load/runtime instead of at the transform boundary. Please re-raise
ValidationErrorhere and only treat non-validation checker failures as skippable.💡 Proposed fix
try: onnx.checker.check_model(model) except onnx.checker.ValidationError as exc: - logger.warning( - "[DLA pipeline] ONNX check reported issues (model may still be usable): %s", exc - ) + logger.error("[DLA pipeline] ONNX validation failed: %s", exc) + raise except Exception as exc: logger.debug( "[DLA pipeline] ONNX in-memory check skipped (model may be too large): %s", exc )🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/make_dla_compatible.py` around lines 162 - 171, The current onnx.checker.check_model call swallows ValidationError and logs a warning but continues; instead, when onnx.checker.ValidationError is caught (from the check_model call), re-raise it so the transform fails fast (do not return/continue). Keep the broader Exception catch for other non-validation issues as skippable and logged with logger.debug; specifically adjust the except onnx.checker.ValidationError branch in the block around onnx.checker.check_model to re-raise the caught ValidationError (while preserving any logging if desired) and only treat general Exception as the skippable case.tests/unit/dla_transforms/test_dla_convert_ops_to_4d.py-161-181 (1)
161-181:⚠️ Potential issue | 🟠 Major | ⚡ Quick winCompare non-floating outputs as well.
These helpers skip value checks for every non-float dtype, so the
ArgMaxcases currently pass as long as the shape matches. That leaves the rewritten indices completely unverified.💡 Proposed fix
def _assert_match(ref: dict, got: dict) -> None: assert ref.keys() == got.keys() for name in ref: a, b = ref[name], got[name] assert a.shape == b.shape, f"{name}: shape {a.shape} vs {b.shape}" if a.dtype.kind == "f": cos = _cosine_similarity(a, b) assert cos >= 1.0 - 1e-5, f"{name}: cosine {cos:.8f}" np.testing.assert_allclose(a, b, rtol=1e-4, atol=1e-4, err_msg=name) + else: + np.testing.assert_array_equal(a, b, err_msg=name) def _assert_values_match(ref: dict, got: dict) -> None: """Element-wise comparison ignoring shape (for ops whose output rank changes).""" assert ref.keys() == got.keys() for name in ref: a, b = ref[name].ravel(), got[name].ravel() assert a.shape == b.shape, f"{name}: element count {a.shape} vs {b.shape}" if a.dtype.kind == "f": cos = _cosine_similarity(a, b) assert cos >= 1.0 - 1e-5, f"{name}: cosine {cos:.8f}" np.testing.assert_allclose(a, b, rtol=1e-4, atol=1e-4, err_msg=name) + else: + np.testing.assert_array_equal(a, b, err_msg=name)🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@tests/unit/dla_transforms/test_dla_convert_ops_to_4d.py` around lines 161 - 181, The helpers _assert_match and _assert_values_match currently only validate floating dtypes' values and skip non-float outputs; update both functions to also compare non-floating arrays element-wise (e.g., using exact equality/assert_array_equal) after the dtype check so integer/bool/index outputs like ArgMax are verified; keep existing cosine/approx checks for float types and apply the element-wise comparison on the flattened arrays in _assert_values_match and on full-shape arrays in _assert_match.
🟡 Minor comments (2)
tests/unit/dla_transforms/test_dla_remove_qdq.py-16-20 (1)
16-20:⚠️ Potential issue | 🟡 Minor | ⚡ Quick winRemove duplicate license header.
Lines 16-20 duplicate the SPDX license header already present at lines 1-14. This appears to be a copy-paste error.
🧹 Proposed fix
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 -# # Unit tests for :mod:`modelopt.onnx.graph_surgery.dla_transforms.dla_remove_qdq`.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@tests/unit/dla_transforms/test_dla_remove_qdq.py` around lines 16 - 20, Remove the duplicated SPDX license header block (the second occurrence containing the SPDX-FileCopyrightText and SPDX-License-Identifier) so only the original header at the top of the file remains; locate the repeated block (the header shown in the diff) in test_dla_remove_qdq.py and delete that duplicate lines 16-20, keeping the rest of the test file unchanged.modelopt/onnx/graph_surgery/dla_transforms/dla_topk.py-43-50 (1)
43-50:⚠️ Potential issue | 🟡 Minor | ⚡ Quick winDocstring/code mismatch: axis condition.
The module docstring (line 47) states the condition is
GatherElements axis == 2, but the code at line 132 checksgather_axis != 1. These are different conditions. Please reconcile the documentation with the implementation.Also applies to: 131-133
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@modelopt/onnx/graph_surgery/dla_transforms/dla_topk.py` around lines 43 - 50, The docstring claims the GatherElements axis must be 2 but the implementation checks gather_axis != 1; reconcile them by making the condition and documentation match: either change the runtime check around gather_axis (the variable used when inspecting the GatherElements node) to require axis == 2, or update the module docstring to state axis == 1. Update the text in the module-level docstring and the conditional that references gather_axis (and any related comments) so they are identical, and run/adjust any unit tests or assertions that rely on the GatherElements axis check.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: abe4dc0a-2f75-4eda-839e-a5b8801873b1
📒 Files selected for processing (44)
modelopt/onnx/graph_surgery/__init__.pymodelopt/onnx/graph_surgery/dla_transforms/_common.pymodelopt/onnx/graph_surgery/dla_transforms/_dla_graph_helpers.pymodelopt/onnx/graph_surgery/dla_transforms/dla_5d_reshape_to_4d.pymodelopt/onnx/graph_surgery/dla_transforms/dla_cast_to_fp32.pymodelopt/onnx/graph_surgery/dla_transforms/dla_constants_to_initializers.pymodelopt/onnx/graph_surgery/dla_transforms/dla_convert_ops_to_4d.pymodelopt/onnx/graph_surgery/dla_transforms/dla_decompose_lstm.pymodelopt/onnx/graph_surgery/dla_transforms/dla_fix_instancenorm_channel_mismatch.pymodelopt/onnx/graph_surgery/dla_transforms/dla_graph_cleanup.pymodelopt/onnx/graph_surgery/dla_transforms/dla_greater.pymodelopt/onnx/graph_surgery/dla_transforms/dla_handle_qdq.pymodelopt/onnx/graph_surgery/dla_transforms/dla_matmul_to_transpose_conv_transpose.pymodelopt/onnx/graph_surgery/dla_transforms/dla_not.pymodelopt/onnx/graph_surgery/dla_transforms/dla_remove_deqlin.pymodelopt/onnx/graph_surgery/dla_transforms/dla_remove_intermediary_squeeze_and_unsqueeze.pymodelopt/onnx/graph_surgery/dla_transforms/dla_remove_qdq.pymodelopt/onnx/graph_surgery/dla_transforms/dla_remove_reshapes.pymodelopt/onnx/graph_surgery/dla_transforms/dla_remove_unused_initializers.pymodelopt/onnx/graph_surgery/dla_transforms/dla_squeeze_unsqueeze_to_reshape.pymodelopt/onnx/graph_surgery/dla_transforms/dla_topk.pymodelopt/onnx/graph_surgery/dla_transforms/dla_unsqueeze.pymodelopt/onnx/graph_surgery/dla_transforms/dla_where.pymodelopt/onnx/graph_surgery/dla_transforms/onnx_dtypes.pymodelopt/onnx/graph_surgery/dq_transpose.pymodelopt/onnx/graph_surgery/encoder_cross_kv.pymodelopt/onnx/graph_surgery/gqa_replacement.pymodelopt/onnx/graph_surgery/make_dla_compatible.pymodelopt/onnx/graph_surgery/utils/dtype_conversion.pytests/unit/dla_transforms/test_dla_5d_reshape_to_4d.pytests/unit/dla_transforms/test_dla_cast_to_fp32.pytests/unit/dla_transforms/test_dla_constants_to_initializers.pytests/unit/dla_transforms/test_dla_convert_ops_to_4d.pytests/unit/dla_transforms/test_dla_decompose_lstm.pytests/unit/dla_transforms/test_dla_fix_instancenorm_channel_mismatch.pytests/unit/dla_transforms/test_dla_graph_cleanup.pytests/unit/dla_transforms/test_dla_greater.pytests/unit/dla_transforms/test_dla_handle_qdq.pytests/unit/dla_transforms/test_dla_matmul_to_transpose_conv_transpose.pytests/unit/dla_transforms/test_dla_not.pytests/unit/dla_transforms/test_dla_remove_deqlin.pytests/unit/dla_transforms/test_dla_remove_qdq.pytests/unit/dla_transforms/test_dla_topk.pytests/unit/dla_transforms/test_dla_where.py
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #1415 +/- ##
==========================================
- Coverage 77.40% 76.62% -0.78%
==========================================
Files 476 496 +20
Lines 51319 54823 +3504
==========================================
+ Hits 39721 42008 +2287
- Misses 11598 12815 +1217
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
1 participant
What does this PR do?
Type of change: new feature
Added graph surgeries to transform models and make them DLA HW compatible.
There is single surgery make_dla_compatible which calls series of sub
Usage
Testing
Added multiple test cases for each sub-surgery and for make_dla_compatible surgery.
Following is the list of test cases added, each containing multiple cases.
tests/unit/dla_transforms/test_dla_5d_reshape_to_4d.py
tests/unit/dla_transforms/test_dla_cast_to_fp32.py
tests/unit/dla_transforms/test_dla_constants_to_initializers.py
tests/unit/dla_transforms/test_dla_convert_ops_to_4d.py
tests/unit/dla_transforms/test_dla_decompose_lstm.py
tests/unit/dla_transforms/test_dla_fix_instancenorm_channel_mismatch.py
tests/unit/dla_transforms/test_dla_graph_cleanup.py
tests/unit/dla_transforms/test_dla_greater.py
tests/unit/dla_transforms/test_dla_handle_qdq.py
tests/unit/dla_transforms/test_dla_matmul_to_transpose_conv_transpose.py
tests/unit/dla_transforms/test_dla_not.py
tests/unit/dla_transforms/test_dla_remove_deqlin.py
tests/unit/dla_transforms/test_dla_remove_qdq.py
tests/unit/dla_transforms/test_dla_topk.py
tests/unit/dla_transforms/test_dla_where.py
Summary by CodeRabbit
Release Notes
New Features
Documentation