tico 0.1.0.dev250923__py3-none-any.whl → 0.1.0.dev250925__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250923"
+__version__ = "0.1.0.dev250925"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/config/v1.py

@@ -24,6 +24,7 @@ class CompileConfigV1(CompileConfigBase):
     convert_lhs_const_mm_to_fc: bool = False
     convert_rhs_const_mm_to_fc: bool = True
     convert_single_batch_lhs_const_bmm_to_fc: bool = False
+    convert_expand_to_slice_cat: bool = False
 
     def get(self, name: str):
         return super().get(name)

tico/experimental/quantization/ptq/wrappers/fairseq/decoder_export_single_step.py

@@ -0,0 +1,234 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# -----------------------------------------------------------------------------
+# This file includes modifications based on fairseq
+#  (https://github.com/facebookresearch/fairseq), originally licensed under
+# the MIT License. See the LICENSE file in the fairseq repository for details.
+# -----------------------------------------------------------------------------
+
+"""
+Q) Why the name "SingleStep"?
+
+Fairseq's decoder already advances one token at a time during generation,
+but the default path is "stateful" and "shape-polymorphic": it owns and
+mutates K/V caches internally, prefix lengths and triangular masks grow with
+the step, and beam reordering updates hidden module state. That's friendly
+for eager execution, but hostile to `torch.export` and many accelerator
+backends.
+
+This export wrapper makes the per-token call truly "single-step" in the
+export sense: "stateless" and "fixed-shape" so every invocation has the
+exact same graph.
+
+Key invariants
+--------------
+• "Stateless": K/V caches come in as explicit inputs and go out as outputs.
+  The module does not store or mutate hidden state.
+• "Static shapes": Query is always [B, 1, C]; encoder features and masks
+  have fixed, predeclared sizes; K/V slots use fixed capacity (unused tail
+  is simply masked/ignored).
+• "External control": Step indexing, cache slot management (append/roll),
+  and beam reordering are handled outside the module.
+• "Prebuilt additive masks": Self-attention masks are provided by the
+  caller (0 for valid, large negative sentinel, e.g. -120, for masked),
+  avoiding data-dependent control flow.
+
+In short: still step-wise like fairseq, but restructured for export—no
+internal state, no data-dependent shapes, no dynamic control flow.
+"""
+
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+
+import tico
+
+# ----- 1) Export wrapper module -------------------------------------------
+class DecoderExportSingleStep(nn.Module):
+    """
+    Export-only single-step decoder module.
+
+    Inputs (example shapes; B=1, H=8, Dh=64, C=512, S=64, Tprev=63):
+      - prev_x:               [B, 1, C]          embedded decoder input for the current step
+      - enc_x:                [S, B, C]          encoder hidden states (fixed-length export input)
+      - enc_pad_additive:     [B, 1, S]          additive float key_padding_mask for enc-dec attn (0 for keep, -120 for pad)
+      - self_attn_mask:       [B, 1, S]          additive float mask for decoder self-attn at this step; pass zeros if unused
+      - prev_self_k_0..L-1:   [B, H, Tprev, Dh]  cached self-attn K per layer
+      - prev_self_v_0..L-1:   [B, H, Tprev, Dh]  cached self-attn V per layer
+
+    Outputs:
+      - x_out:                [B, 1, C]          new decoder features at the current step
+      - new_k_0..L-1:         [H, B, Dh]         per-layer new K (single-timestep; time dim squeezed)
+      - new_v_0..L-1:         [H, B, Dh]         per-layer new V (single-timestep; time dim squeezed)
+
+    Notes:
+      • We keep masks/additive semantics externally to avoid any mask-building inside the graph.
+      • We reshape the new K/V from [B,H,1,Dh] -> [H,B,Dh] to match the requested output spec (8,1,64).
+    """
+
+    def __init__(self, decoder: nn.Module):
+        super().__init__()
+        self.decoder = decoder
+        # Cache common meta for assertions
+        self.num_layers = len(getattr(decoder, "layers"))
+        # Infer heads/head_dim from the wrapped self_attn of layer 0
+        any_layer = getattr(decoder.layers[0], "wrapped", decoder.layers[0])  # type: ignore[index]
+        mha = getattr(any_layer, "self_attn", None)
+        assert mha is not None, "Decoder layer must expose self_attn"
+        self.num_heads = int(mha.num_heads)
+        self.head_dim = int(mha.head_dim)
+        # Embed dim (C)
+        self.embed_dim = int(getattr(decoder, "embed_dim"))
+
+    def forward(
+        self,
+        prev_x: torch.Tensor,  # [B,1,C]
+        enc_x: torch.Tensor,  # [S,B,C]
+        enc_pad_additive: torch.Tensor,  # [B,1,S]
+        *kv_args: torch.Tensor,  # prev_k_0..L-1, prev_v_0..L-1 (total 2L tensors)
+        self_attn_mask: torch.Tensor,  # [B,1,S] (or zeros)
+    ):
+        L = self.num_layers
+        H = self.num_heads
+        Dh = self.head_dim
+        B, one, C = prev_x.shape
+        S, B2, C2 = enc_x.shape
+        assert (
+            one == 1 and C == self.embed_dim and B == B2 and C2 == C
+        ), "Shape mismatch in prev_x/enc_x"
+        assert len(kv_args) == 2 * L, f"Expected {2*L} KV tensors, got {len(kv_args)}"
+
+        # Unpack previous self-attn caches
+        prev_k_list: List[torch.Tensor] = list()  # each [B,H,Tprev,Dh]
+        prev_v_list: List[torch.Tensor] = list()  # each [B,H,Tprev,Dh]
+        for i in range(L):
+            prev_k_list.append(kv_args[2 * i])
+            prev_v_list.append(kv_args[2 * i + 1])
+        for i in range(L):
+            assert (
+                prev_k_list[i].dim() == 4 and prev_v_list[i].dim() == 4
+            ), "KV must be [B,H,Tprev,Dh]"
+            assert (
+                prev_k_list[i].shape[0] == B
+                and prev_k_list[i].shape[1] == H
+                and prev_k_list[i].shape[3] == Dh
+            )
+
+        # Call decoder's external single-step path
+        # Returns:
+        #   x_step: [B,1,C]
+        #   newk/newv: lists of length L, each [B*H,1,Dh]
+        x_step, newk_list, newv_list = self.decoder.forward_external_step(  # type: ignore[operator]
+            prev_output_x=prev_x,
+            encoder_out_x=enc_x,
+            encoder_padding_mask=enc_pad_additive,
+            self_attn_mask=self_attn_mask,
+            prev_self_k_list=prev_k_list,
+            prev_self_v_list=prev_v_list,
+        )
+
+        out_tensors: List[torch.Tensor] = [
+            x_step
+        ]  # first output is the new decoder features
+        for i in range(L):
+            nk = newk_list[i]  # [B*H, Tnew, Dh]
+            nv = newv_list[i]  # [B*H, Tnew, Dh]
+            out_tensors.append(nk)
+            out_tensors.append(nv)
+
+        # Return tuple: (x_step, new_k_0, new_v_0, new_k_1, new_v_1, ..., new_k_{L-1}, new_v_{L-1})
+        return tuple(out_tensors)
+
+
+# ----- 2) Example inputs (B=1, S=64, H=8, Dh=64, C=512, L=4) ---------------
+def make_example_inputs(*, L=4, B=1, S=64, H=8, Dh=64, C=512, Tprev=63, device="cpu"):
+    """
+    Build example tensors that match the export I/O spec.
+    Shapes follow the request:
+      prev_x:             [1,1,512]
+      enc_x:              [64,1,512]
+      enc_pad_additive:   [1,1,64]    (additive float; zeros -> keep)
+      prev_k_i / prev_v_i (for i in 0..L-1): [1,8,63,64]
+      self_attn_mask:     [1,1,64]    (additive float; zeros -> keep)
+    """
+    g = torch.Generator(device=device).manual_seed(0)
+
+    prev_x = torch.randn(B, 1, C, device=device, dtype=torch.float32, generator=g)
+    enc_x = torch.randn(S, B, C, device=device, dtype=torch.float32, generator=g)
+
+    # Additive masks (0 for allowed, -120 for masked)
+    enc_pad_additive = torch.full((B, 1, S), float(-120), device=device)
+    self_attn_mask = torch.full((B, 1, S), float(-120), device=device)
+    enc_pad_additive[0, :27] = 0  # 27 is a random example.
+    self_attn_mask[0, :27] = 0  # 27 is a random example.
+
+    # Previous self-attn caches for each layer
+    prev_k_list = []
+    prev_v_list = []
+    for _ in range(L):
+        prev_k = torch.randn(
+            B, H, Tprev, Dh, device=device, dtype=torch.float32, generator=g
+        )
+        prev_v = torch.randn(
+            B, H, Tprev, Dh, device=device, dtype=torch.float32, generator=g
+        )
+        prev_k_list.append(prev_k)
+        prev_v_list.append(prev_v)
+
+    # Pack inputs as the export function will expect:
+    # (prev_x, enc_x, enc_pad_additive, self_attn_mask, prev_k_0..L-1, prev_v_0..L-1)
+    example_args: Tuple[torch.Tensor, ...] = (
+        prev_x,
+        enc_x,
+        enc_pad_additive,
+        *prev_k_list,
+        *prev_v_list,
+    )
+    example_kwargs = {"self_attn_mask": self_attn_mask}
+    return example_args, example_kwargs
+
+
+# ----- 3) Export driver -----------------------------------------------------
+def export_decoder_single_step(translator, *, save_path="decoder_step_export.circle"):
+    """
+    Wrap the QuantFairseqDecoder into the export-friendly single-step module
+    and export with torch.export.export using example inputs.
+    """
+    # Grab the wrapped decoder
+    dec = translator.models[
+        0
+    ].decoder  # assumed QuantFairseqDecoder with forward_external_step
+    # Build export wrapper
+    wrapper = DecoderExportSingleStep(decoder=dec).eval()
+
+    # Example inputs (L inferred from wrapper/decoder)
+    L = wrapper.num_layers
+    H = wrapper.num_heads
+    Dh = wrapper.head_dim
+    C = wrapper.embed_dim
+    example_inputs, example_kwargs = make_example_inputs(L=L, H=H, Dh=Dh, C=C)
+
+    # Export circle (no dynamism assumed; shapes are fixed for export)
+    cm = tico.convert(
+        wrapper,
+        args=example_inputs,
+        kwargs=example_kwargs,
+        strict=True,  # fail if something cannot be captured
+    )
+
+    # Save .pte
+    cm.save(save_path)
+    print(f"Saved decoder single-step export to: {save_path}")

tico/passes/convert_expand_to_slice_cat.py

@@ -0,0 +1,153 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+import torch
+from torch.export import ExportedProgram
+
+from tico.passes import ops
+from tico.serialize.circle_mapping import extract_shape
+from tico.utils import logging
+from tico.utils.graph import create_node
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_graph_diff_on_pass
+from tico.utils.utils import is_target_node
+from tico.utils.validate_args_kwargs import ExpandArgs, ReshapeArgs
+
+
+@trace_graph_diff_on_pass
+class ConvertExpandToSliceCat(PassBase):
+    """
+    This pass replaces `aten.reshape` + `aten.expand` pattern by rewriting it using
+    a series of `aten.slice` and `aten.cat` operations.
+
+    This pass is specialized for expand of KVCache.
+    - Expects (batch, num_key_value_heads, seq_len, head_dim) as input shape of reshape
+    """
+
+    def __init__(self, enabled: bool = False):
+        super().__init__()
+        self.enabled = enabled
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        if not self.enabled:
+            return PassResult(False)
+
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+
+        # This pass handles expand on EXPAND_DIM only
+        CAT_DIM = 1
+        EXPAND_DIM = 2
+
+        for node in graph.nodes:
+            if not isinstance(node, torch.fx.Node) or not is_target_node(
+                node, ops.aten.reshape
+            ):
+                continue
+
+            post_reshape = node
+            post_reshape_args = ReshapeArgs(*post_reshape.args, **post_reshape.kwargs)
+            post_reshape_input = post_reshape_args.input
+
+            if not isinstance(post_reshape_input, torch.fx.Node) or not is_target_node(
+                post_reshape_input, ops.aten.expand
+            ):
+                continue
+
+            expand = post_reshape_input
+            expand_args = ExpandArgs(*expand.args, **expand.kwargs)
+            expand_input = expand_args.input
+            expand_shape = extract_shape(expand)
+
+            if not isinstance(expand_input, torch.fx.Node) or not is_target_node(
+                expand_input, ops.aten.reshape
+            ):
+                continue
+
+            pre_reshape = expand_input
+            pre_reshape_args = ReshapeArgs(*pre_reshape.args, **pre_reshape.kwargs)
+            pre_reshape_input = pre_reshape_args.input
+            pre_reshape_shape = extract_shape(pre_reshape)
+
+            if pre_reshape_shape[EXPAND_DIM] != 1:
+                continue
+
+            reshape_input_shape = extract_shape(pre_reshape_input)
+
+            if len(expand_shape) != len(pre_reshape_shape):
+                continue
+
+            # Ensure all dimensions *except* at EXPAND_DIM are identical.
+            if not (
+                expand_shape[:EXPAND_DIM] == pre_reshape_shape[:EXPAND_DIM]
+                and expand_shape[EXPAND_DIM + 1 :]
+                == pre_reshape_shape[EXPAND_DIM + 1 :]
+            ):
+                continue
+
+            # Ensure the expansion dimension is a clean multiple.
+            if expand_shape[EXPAND_DIM] % pre_reshape_shape[EXPAND_DIM] != 0:
+                continue
+
+            expand_ratio = expand_shape[EXPAND_DIM] // pre_reshape_shape[EXPAND_DIM]
+
+            if expand_ratio <= 1:
+                continue
+
+            cat_nodes = []
+
+            for i in range(reshape_input_shape[CAT_DIM]):
+                with graph.inserting_before(expand):
+                    slice_copy_args = (pre_reshape_input, CAT_DIM, i, i + 1, 1)
+                    slice_node = create_node(
+                        graph,
+                        torch.ops.aten.slice.Tensor,
+                        args=slice_copy_args,
+                        origin=expand,
+                    )
+                with graph.inserting_after(slice_node):
+                    cat_args = ([slice_node] * expand_ratio, CAT_DIM)
+                    cat_node = create_node(
+                        graph,
+                        torch.ops.aten.cat.default,
+                        args=cat_args,
+                        origin=expand,
+                    )
+                    cat_nodes.append(cat_node)
+
+            with graph.inserting_after(expand):
+                cat_args = (cat_nodes, CAT_DIM)
+                cat_node = create_node(
+                    graph,
+                    torch.ops.aten.cat.default,
+                    args=cat_args,
+                    origin=expand,
+                )
+                expand.replace_all_uses_with(cat_node)
+
+            modified = True
+            logger.debug(f"{expand.name} is replaced with {cat_node.name} operators")
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/ops.py

@@ -69,7 +69,6 @@ class AtenOps:
             torch.ops.aten.unsqueeze_copy.default,
         ]
         self.view = [
-            torch.ops.aten.view,
             torch.ops.aten.view.default,
             torch.ops.aten.view_copy.default,
         ]

tico/utils/convert.py

@@ -39,6 +39,7 @@ from tico.passes.cast_clamp_mixed_type_args import CastClampMixedTypeArgs
 from tico.passes.cast_mixed_type_args import CastMixedTypeArgs
 from tico.passes.const_prop_pass import ConstPropPass
 from tico.passes.convert_conv1d_to_conv2d import ConvertConv1dToConv2d
+from tico.passes.convert_expand_to_slice_cat import ConvertExpandToSliceCat
 from tico.passes.convert_layout_op_to_reshape import ConvertLayoutOpToReshape
 from tico.passes.convert_matmul_to_linear import ConvertMatmulToLinear
 from tico.passes.convert_repeat_to_expand_copy import ConvertRepeatToExpandCopy
@@ -250,6 +251,7 @@ def convert_exported_module_to_circle(
             ConstPropPass(),
             SegmentIndexSelectConst(),
             LegalizeCausalMaskValue(enabled=config.get("legalize_causal_mask_value")),
+            ConvertExpandToSliceCat(enabled=config.get("convert_expand_to_slice_cat")),
             ConvertMatmulToLinear(
                 enable_lhs_const=config.get("convert_lhs_const_mm_to_fc"),
                 enable_rhs_const=config.get("convert_rhs_const_mm_to_fc"),

{tico-0.1.0.dev250923.dist-info → tico-0.1.0.dev250925.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250923
+Version: 0.1.0.dev250925
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250923.dist-info → tico-0.1.0.dev250925.dist-info}/RECORD

@@ -1,9 +1,9 @@
-tico/__init__.py,sha256=SQIp1AUtHO6svZTLOBdHKzr7ss9Z3l9yh8cyySsGZOo,1883
+tico/__init__.py,sha256=Dh35NwoATHo8LtQfaClJ4ecMmZjFjQAmdKJ9nJvnQb0,1883
 tico/pt2_to_circle.py,sha256=gu3MD4Iqc0zMZcCZ2IT8oGbyj21CTSbT3Rgd9s2B_9A,2767
 tico/config/__init__.py,sha256=xZzCXjZ84qE-CsBi-dfaL05bqpQ3stKKfTXhnrJRyVs,142
 tico/config/base.py,sha256=q5xMqGxTUZs4mFqt5c7i_y9U00fYgdMGl9nUqIVMlCo,1248
 tico/config/factory.py,sha256=il0zqB6Lm5NX2LnG-TUhmiP9vVeZ_3TucJMorVZIodY,1324
-tico/config/v1.py,sha256=lEyKemeKGrJ0bA5w-LPkMWVlnAiJRDm9mM48TJle-e4,1296
+tico/config/v1.py,sha256=uB5d39fkmuBACwjBVGtdWb_HGXfXsvmw6nw64xZcC-8,1342
 tico/experimental/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/__init__.py,sha256=IaJPZegVJp0P3luutBo907Kp5sOJensE1Mm-XBG_jBs,122
 tico/experimental/quantization/public_interface.py,sha256=TGo3bTapwLA8KpsoEwBhuzI0LQUO6y3-sUM1VZvkLo8,4220
@@ -91,6 +91,7 @@ tico/experimental/quantization/ptq/wrappers/quant_elementwise.py,sha256=LhEoobfv
 tico/experimental/quantization/ptq/wrappers/quant_module_base.py,sha256=vkcDos_knGSS29rIZuEIWkAJLHrENbGz8nCH2-iara8,5969
 tico/experimental/quantization/ptq/wrappers/registry.py,sha256=OVO5nev6J8Br9zsIX-Ut7ZgWzA9f_jk0Np9bGioXgQM,5171
 tico/experimental/quantization/ptq/wrappers/fairseq/__init__.py,sha256=Mc8FLd9DusyB_IT1vk1OYrRkngOYnYd05IvtA9ORVQc,160
+tico/experimental/quantization/ptq/wrappers/fairseq/decoder_export_single_step.py,sha256=d7ZieKiSbZ2ffkaLYMg2PJl1OyAxkKjB3OHKB4poxJs,9796
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_decoder.py,sha256=CILYvxPhW7xLkroWW_hunQBGAYGexLqnPnO5xmMnK-E,17877
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_decoder_layer.py,sha256=JT79shxOhDtRFgm8jrrN6HKvyVotiytLjMjAxX-Cztg,20416
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_encoder.py,sha256=r9DPUAbL2KRJ8zpMJ39Y9n6Oe79nte-mFcdjG2qEP-w,13809
@@ -113,6 +114,7 @@ tico/passes/cast_clamp_mixed_type_args.py,sha256=m3_HpXLywWmWERfE5lM5PgvjBod7C4B
 tico/passes/cast_mixed_type_args.py,sha256=Wd3sCDKJZwdb8GiMWKljm8X5CLFRd8eCz-dmWks15Hc,7763
 tico/passes/const_prop_pass.py,sha256=hDxGgJNiRjsgOArdaoeAOcOOA-nKBvA1W1zcMZQA5yg,11531
 tico/passes/convert_conv1d_to_conv2d.py,sha256=ktS3h158y9rg1sQiW8BZZbflV_dk_UdjBPQnuiOKyzg,5303
+tico/passes/convert_expand_to_slice_cat.py,sha256=Fa6b5pqiQNq-QBiEC0e3WkQYf2UEhMgzSTIt4hlzdjc,5470
 tico/passes/convert_layout_op_to_reshape.py,sha256=sCAFjkmVtiKjvDQSAgnjNBHl3_hWXJZElGDXQiTH-7s,2963
 tico/passes/convert_matmul_to_linear.py,sha256=WATtsHk_GzsU0HYovc3UMyEj8ApF2qLbInAsNlQj0nE,9759
 tico/passes/convert_repeat_to_expand_copy.py,sha256=JbtFTmWyfJS2SSd_higP1IEhQeh7wHdN5dmTbbiFVCs,3237
@@ -134,7 +136,7 @@ tico/passes/lower_pow2_to_mul.py,sha256=nfJXa9ZTZMiLg6ownSyvkM4KF2z9tZW34Q3CCWI_
 tico/passes/lower_to_resize_nearest_neighbor.py,sha256=gbrvTmWSXDPdJ1XJtWGI5mo-uEiauXEG3ELwbKYVPLI,9013
 tico/passes/lower_to_slice.py,sha256=OzlFzK3lBYyYwC3WThsWd94Ob4JINIJF8UaLAtnumzU,7262
 tico/passes/merge_consecutive_cat.py,sha256=ayZNLDA1DFM7Fxxi2Dmk1CujkgUuaVCH1rhQgLrvvOQ,2701
-tico/passes/ops.py,sha256=cSj3Sk2x2cOE9b8oU5pmSa_rHr-iX2lORzu3N_UHMSQ,2967
+tico/passes/ops.py,sha256=7IGRnxIJl-nLO4huVk_mgBfD4VGUNQRyeuM8K1L2u1U,2934
 tico/passes/remove_nop.py,sha256=Hf91p_EJAOC6DyWNthash0_UWtEcNc_M7znamQfYQ5Y,2686
 tico/passes/remove_redundant_assert_nodes.py,sha256=rYbTCyuNIXIC-2NreHKBVCuaSUkEQvB_iSRzb26P_EA,1821
 tico/passes/remove_redundant_expand.py,sha256=8yhlMnbog-T9gIK6LKIU0tu0__gfhZzO36g_fJIVVP4,2162
@@ -236,7 +238,7 @@ tico/serialize/operators/utils.py,sha256=lXGpEJW1h8U_-gfc6EWjvvSiq3yJ9P-v1v3EMRT
 tico/serialize/operators/adapters/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/serialize/operators/adapters/llama_rmsnorm.py,sha256=6t3dhfNpR03eIjsmhymF2JKd6lCf7PvInqMf77c_BOE,1139
 tico/utils/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/utils/convert.py,sha256=XbogVXO-QS0UTFNvEDyADvhCp87kTUpGAUalN8I8eRQ,13645
+tico/utils/convert.py,sha256=10YufXpuqHz274ACUb1_F5594uClUFhBEh8SY6gYp7w,13809
 tico/utils/define.py,sha256=Ypgp7YffM4pgPl4Zh6TmogSn1OxGBMRw_e09qYGflZk,1467
 tico/utils/diff_graph.py,sha256=_eDGGPDPYQD4b--MXX0DLoVgSt_wLfNPt47UlolLLR4,5272
 tico/utils/dtype.py,sha256=L5Qb7qgbt0eQ5frUTvHYrRtTJb1dg4-JNEopcxCNg1U,1389
@@ -260,9 +262,9 @@ tico/utils/mx/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/utils/mx/elemwise_ops.py,sha256=V6glyAHsVR1joqpsgnNytatCD_ew92xNWZ19UFDoMTA,10281
 tico/utils/mx/formats.py,sha256=uzNWyu-1onUlwQfX5cZ6fZSUfHMRqorper7_T1k3jfk,3404
 tico/utils/mx/mx_ops.py,sha256=RcfUTYVi-wilGB2sC35OeARdwDqnixv7dG5iyZ-fQT8,8555
-tico-0.1.0.dev250923.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250923.dist-info/METADATA,sha256=YFowPhlSZB_-sCJe9E7tmXhQdMo0EWQ-H9fPL57i6SQ,8450
-tico-0.1.0.dev250923.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250923.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250923.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250923.dist-info/RECORD,,
+tico-0.1.0.dev250925.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250925.dist-info/METADATA,sha256=LReZIvZizVYMd9esGIdPgEiaO250UU59SBEoEbAe0rM,8450
+tico-0.1.0.dev250925.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250925.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250925.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250925.dist-info/RECORD,,