tico 0.1.0.dev250831__py3-none-any.whl → 0.1.0.dev250902__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250831"
+__version__ = "0.1.0.dev250902"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/ptq/examples/compare_ppl.py

@@ -0,0 +1,121 @@
+# 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.
+
+# =============================================================================
+# QUICK PTQ WORKFLOW (OPTIONAL FP32 BASELINE)
+# -----------------------------------------------------------------------------
+# Toggle RUN_FP to choose between:
+#   • FP32 perplexity measurement only, OR
+#   • Full post-training UINT-8 flow (wrap → calibrate → eval).
+# =============================================================================
+
+import torch
+import tqdm
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.utils import perplexity
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+
+# -------------------------------------------------------------------------
+# 0. Global configuration
+# -------------------------------------------------------------------------
+MODEL_NAME = "meta-llama/Meta-Llama-3-1B"
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+STRIDE = 512  # sliding-window stride for perplexity
+RUN_FP = True  # set False → run UINT-8 path
+
+# Token-budget presets for activation calibration
+TOKENS: dict[str, int] = {
+    # Smoke test (<1 min turnaround on CPU/GPU)
+    "debug": 2_000,  # ≈16 × 128-seq batches
+    # Good default for 1-7B models (≲3 % ppl delta)
+    "baseline": 50_000,
+    # Production / 4-bit observer smoothing
+    "production": 200_000,
+}
+CALIB_TOKENS = TOKENS["baseline"]
+print(f"Calibrating with {CALIB_TOKENS:,} tokens.\n")
+
+# -------------------------------------------------------------------------
+# 1. Load model
+# -------------------------------------------------------------------------
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+
+if RUN_FP:
+    # -- FP32 baseline ------------------------------------------------------
+    print("Loading FP32 model …")
+    fp_model = AutoModelForCausalLM.from_pretrained(MODEL_NAME).to(DEVICE).eval()
+    fp_model.config.use_cache = False
+else:
+    # -- UINT-8 pipeline -----------------------------------------------------
+    print("Creating UINT-8 clone …")
+    uint8_model = AutoModelForCausalLM.from_pretrained(MODEL_NAME).to(DEVICE).eval()
+    uint8_model.config.use_cache = False
+
+    # ---------------------------------------------------------------------
+    # 2. Wrap every Transformer layer with PTQWrapper
+    # ---------------------------------------------------------------------
+    qcfg = QuantConfig()  # all-uint8 defaults
+
+    wrapped_layers = torch.nn.ModuleList()
+    for idx, layer in enumerate(uint8_model.model.layers):
+        layer_cfg = qcfg.child(f"layer{idx}")
+        wrapped_layers.append(PTQWrapper(layer, qcfg=layer_cfg))
+    uint8_model.model.layers = wrapped_layers
+
+    # ---------------------------------------------------------------------
+    # 3. Single-pass activation calibration
+    # ---------------------------------------------------------------------
+    print("Calibrating UINT-8 observers …")
+    calib_txt = " ".join(
+        load_dataset("wikitext", "wikitext-2-raw-v1", split="train")["text"]
+    )[:CALIB_TOKENS]
+    ids = tokenizer(calib_txt, return_tensors="pt").input_ids.to(DEVICE)
+
+    # (a) switch every QuantModuleBase to CALIB mode
+    for l in uint8_model.model.layers:
+        l.enable_calibration()
+
+    # (b) run inference to collect ranges
+    with torch.no_grad():
+        for i in tqdm.trange(0, ids.size(1) - 1, STRIDE, desc="Calibration"):
+            uint8_model(ids[:, i : i + STRIDE])
+
+    # (c) freeze (scale, zero-point)
+    for l in uint8_model.model.layers:
+        l.freeze_qparams()
+
+# -------------------------------------------------------------------------
+# 4. Evaluate perplexity on Wikitext-2
+# -------------------------------------------------------------------------
+print("\nCalculating perplexities …")
+test_ds = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
+enc = tokenizer("\n\n".join(test_ds["text"]), return_tensors="pt")
+
+if RUN_FP:
+    ppl_fp = perplexity(fp_model, enc, DEVICE, stride=STRIDE)
+else:
+    ppl_int8 = perplexity(uint8_model, enc, DEVICE, stride=STRIDE)
+
+# -------------------------------------------------------------------------
+# 5. Report
+# -------------------------------------------------------------------------
+print("\n┌── Wikitext-2 test perplexity ─────────────")
+if RUN_FP:
+    print(f"│ FP32  : {ppl_fp:8.2f}")
+else:
+    print(f"│ UINT-8 : {ppl_int8:8.2f}")
+print("└───────────────────────────────────────────")

tico/experimental/quantization/ptq/examples/quantize_llama_decoder_layer.py

@@ -0,0 +1,124 @@
+# 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.
+
+# =============================================================================
+# POST-TRAINING QUANTIZATION EXAMPLE — Llama Decoder Layer (Self-Attn + MLP)
+# -----------------------------------------------------------------------------
+# This demo shows how to:
+#   1. Replace a single FP32 `LlamaDecoderLayer` with `QuantLlamaDecoderLayer`.
+#   2. Collect activation statistics in one calibration sweep.
+#   3. Freeze scales / zero-points and switch to INT-simulation mode.
+#   4. Compare INT-8 vs FP32 outputs with a quick mean-absolute-diff check.
+#   5. Export the calibrated, quantized block to a Circle model.
+# -----------------------------------------------------------------------------
+# Style / layout is kept identical to the `quantize_llama_attn.py` and
+# `quantize_llama_mlp.py` examples for easy side-by-side reading.
+# =============================================================================
+
+import pathlib
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from tico.experimental.quantization.evaluation.metric import compute_peir
+from tico.experimental.quantization.evaluation.utils import plot_two_outputs
+from tico.experimental.quantization.ptq.mode import Mode
+from tico.experimental.quantization.ptq.wrappers.llama.quant_decoder_layer import (
+    QuantLlamaDecoderLayer,
+)
+from tico.utils.utils import SuppressWarning
+
+MODEL_NAME = "Maykeye/TinyLLama-v0"
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+
+model.eval()  # disable dropout, etc.
+rotary = model.model.rotary_emb  # RoPE helper
+
+# -------------------------------------------------------------------------
+# 1. Swap in the quant wrapper
+# -------------------------------------------------------------------------
+fp32_layer = model.model.layers[0]  # keep a reference for diff check
+model.model.layers[0] = QuantLlamaDecoderLayer(
+    fp32_layer
+)  # PTQWrapper(fp32_layer) is also fine
+model.eval()
+
+qlayer = model.model.layers[0]  # alias for brevity
+
+# -------------------------------------------------------------------------
+# 2. Single-pass calibration (gather activation ranges)
+# -------------------------------------------------------------------------
+PROMPTS = [
+    "The quick brown fox jumps over the lazy dog.",
+    "In 2025, AI systems accelerated hardware-software co-design at scale.",
+    "양자화는 왜 어려울까? 분포, 길이, 마스크가 관건이다.",
+    "今日はいい天気ですね。ところでRoPE角度は長さに依存します。",
+    "def quicksort(arr):\n    if len(arr) <= 1: return arr\n    ...",
+    "Prices rose 3.14% — see Figure 2; emails: foo@bar.com!",
+]
+
+with torch.no_grad():
+    qlayer.enable_calibration()
+    for prompt in PROMPTS:
+        ids = tokenizer(prompt, return_tensors="pt")
+        hidden = model.model.embed_tokens(ids["input_ids"])
+        pos = rotary(hidden, ids["input_ids"])  # (cos, sin) tuple
+        S = pos[0].shape[1]
+        attn_mask = torch.zeros(1, 1, S, S)  # causal-mask placeholder
+        _ = qlayer(hidden, attention_mask=attn_mask, position_embeddings=pos)
+    qlayer.freeze_qparams()
+
+assert qlayer._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick INT-sim vs FP32 sanity check
+# -------------------------------------------------------------------------
+ids = tokenizer("check", return_tensors="pt")
+hidden = model.model.embed_tokens(ids["input_ids"])
+pos = rotary(hidden, ids["input_ids"])
+S = pos[0].shape[1]
+attn_mask = torch.zeros(1, 1, S, S)
+
+with torch.no_grad():
+    int8_out = qlayer(hidden, attention_mask=attn_mask, position_embeddings=pos)
+    int8 = int8_out[0] if isinstance(int8_out, tuple) else int8_out
+    fp32_out = fp32_layer(hidden, attention_mask=attn_mask, position_embeddings=pos)
+    fp32 = fp32_out[0] if isinstance(fp32_out, tuple) else fp32_out
+
+print("┌───────────── Quantization Error Summary ─────────────")
+print(f"│ Mean |diff|: {(int8 - fp32).abs().mean().item():.6f}")
+print(f"│ PEIR       : {compute_peir(fp32, int8) * 100:.6f} %")
+print("└──────────────────────────────────────────────────────")
+print(plot_two_outputs(fp32, int8))
+
+# -------------------------------------------------------------------------
+# 4. Export the calibrated layer to Circle
+# -------------------------------------------------------------------------
+import tico
+
+save_path = pathlib.Path("decoder_layer.q.circle")
+B, S, D = 1, 4, model.config.hidden_size
+example_hidden = torch.randn(B, S, D)
+example_pos = rotary(example_hidden, torch.arange(S)[None, :])
+attn_mask = torch.zeros(1, 1, S, S)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(
+        qlayer, (example_hidden, attn_mask), {"position_embeddings": example_pos}
+    )
+# Note that the model is not fully quantized.
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")

tico/experimental/quantization/ptq/utils/__init__.py

@@ -1,5 +1,7 @@
+from tico.experimental.quantization.ptq.utils.metrics import perplexity
 from tico.experimental.quantization.ptq.utils.reduce_utils import channelwise_minmax
 
 __all__ = [
     "channelwise_minmax",
+    "perplexity",
 ]

tico/experimental/quantization/ptq/utils/metrics.py

@@ -0,0 +1,123 @@
+# 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 Optional
+
+import torch
+import tqdm
+
+
+def perplexity(
+    model: torch.nn.Module,
+    encodings: torch.Tensor,
+    device: torch.device | str,
+    *,
+    max_length: Optional[int] = None,
+    stride: int = 512,
+    ignore_index: int | None = -100,
+    show_progress: bool = True,
+) -> float:
+    """
+    Compute perplexity (PPL) using a "strided sliding-window"
+     evaluation strategy.
+
+    The function:
+    1. Splits the token sequence into overlapping windows of length
+       `max_length` (model context size).
+    2. Masks tokens that were already scored in previous windows
+       (`labels == -100`), so each token's negative log-likelihood (NLL)
+       is counted EXACTLY once.
+    3. Aggregates token-wise NLL to return corpus-level PPL.
+
+    Parameters
+    ----------
+    model : torch.nn.Module
+        Causal LM loaded in evaluation mode (`model.eval()`).
+    encodings : torch.Tensor | transformers.BatchEncoding
+        Tokenised corpus.  If a `BatchEncoding` is passed, its
+        `.input_ids` field is used.  Shape must be `(1, seq_len)`.
+    device : torch.device | str
+        CUDA or CPU device on which to run evaluation.
+    max_length : int, optional
+        Context window size.  Defaults to `model.config.max_position_embeddings`.
+    stride : int, default = 512
+        Step size by which the sliding window advances.  Must satisfy
+        `1 ≤ stride ≤ max_length`.
+    ignore_index : int, default = -100
+        Label value to ignore in loss computation. This should match
+        the `ignore_index` used by the model's internal
+        `CrossEntropyLoss`. For Hugging Face causal LMs, the
+        convention is `-100`.
+    show_progress : bool, default = True
+        If True, displays a tqdm progess bar while evaluating.
+
+    Returns
+    -------
+    float
+        Corpus-level perplexity.
+    """
+    # -------- input preparation -------- #
+    try:
+        # transformers.BatchEncoding has `input_ids`
+        input_ids_full = encodings.input_ids  # type: ignore[attr-defined]
+    except AttributeError:  # already a tensor
+        input_ids_full = encodings
+    assert isinstance(input_ids_full, torch.Tensor)
+    input_ids_full = input_ids_full.to(device)
+
+    if max_length is None:
+        assert hasattr(model, "config")
+        assert hasattr(model.config, "max_position_embeddings")
+        assert isinstance(model.config.max_position_embeddings, int)
+        max_length = model.config.max_position_embeddings
+    assert max_length is not None
+    assert (
+        1 <= stride <= max_length
+    ), f"stride ({stride}) must be in [1, max_length ({max_length})]"
+
+    seq_len = input_ids_full.size(1)
+    nll_sum = 0.0
+    n_tokens = 0
+    prev_end = 0
+
+    # -------- main loop -------- #
+    for begin in tqdm.trange(0, seq_len, stride, desc="PPL", disable=not show_progress):
+        end = min(begin + max_length, seq_len)
+        trg_len = end - prev_end  # fresh tokens in this window
+
+        input_ids = input_ids_full[:, begin:end]
+        target_ids = input_ids.clone()
+        target_ids[:, :-trg_len] = ignore_index  # mask previously-scored tokens
+
+        with torch.no_grad():
+            outputs = model(input_ids, labels=target_ids)
+            # loss is already averaged over non-masked labels
+            neg_log_likelihood = outputs.loss
+
+        # exact number of labels that contributed to loss
+        loss_tokens = (target_ids[:, 1:] != ignore_index).sum().item()
+        nll_sum += neg_log_likelihood * loss_tokens
+        n_tokens += int(loss_tokens)
+
+        prev_end = end
+        if end == seq_len:
+            break
+
+    avg_nll: float | torch.Tensor = nll_sum / n_tokens
+    if not isinstance(avg_nll, torch.Tensor):
+        avg_nll = torch.tensor(avg_nll)
+    assert isinstance(avg_nll, torch.Tensor)
+    ppl = torch.exp(avg_nll)
+
+    return ppl.item()

tico/experimental/quantization/ptq/wrappers/llama/__init__.py

@@ -1,6 +1,9 @@
 from tico.experimental.quantization.ptq.wrappers.llama.quant_attn import (
     QuantLlamaAttention,
 )
+from tico.experimental.quantization.ptq.wrappers.llama.quant_decoder_layer import (
+    QuantLlamaDecoderLayer,
+)
 from tico.experimental.quantization.ptq.wrappers.llama.quant_mlp import QuantLlamaMLP
 
-__all__ = ["QuantLlamaAttention", "QuantLlamaMLP"]
+__all__ = ["QuantLlamaAttention", "QuantLlamaDecoderLayer", "QuantLlamaMLP"]

tico/experimental/quantization/ptq/wrappers/llama/quant_decoder_layer.py

@@ -0,0 +1,168 @@
+# 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 Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.llama.quant_attn import (
+    QuantLlamaAttention,
+)
+from tico.experimental.quantization.ptq.wrappers.llama.quant_mlp import QuantLlamaMLP
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import try_register
+
+
+@try_register("transformers.models.llama.modeling_llama.LlamaDecoderLayer")
+class QuantLlamaDecoderLayer(QuantModuleBase):
+    """
+    Quant-aware drop-in replacement for HF `LlamaDecoderLayer`.
+    Signature and return-value are identical to the original.
+
+    ▸ Attention & MLP blocks are replaced by their quantized counterparts
+    ▸ LayerNorms remain FP32 (no fake-quant)
+    ▸ A "static" causal mask is pre-built in `__init__` to avoid
+      dynamic boolean-to-float casts inside `forward`.
+
+    Notes on the causal mask
+    ------------------------
+    Building a boolean mask "inside" `forward` would introduce
+    non-deterministic dynamic ops that an integer-only accelerator cannot
+    fuse easily.  Therefore we:
+
+    1. Pre-compute a full upper-triangular mask of size
+       `[1, 1, max_seq, max_seq]` in `__init__`.
+    2. In `forward`, if the caller passes `attention_mask=None`, we
+       slice the pre-computed template to the current sequence length.
+    """
+
+    def __init__(
+        self,
+        fp_layer: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+        return_type: Optional[str] = None,
+    ):
+        """
+        Q) Why do we need `return_type`?
+        A) Different versions of `transformers` wrap the decoder output in
+            different containers: a plain Tensor or a tuple.
+        """
+        self.return_type = return_type
+        if self.return_type is None:
+            import transformers
+
+            v = tuple(map(int, transformers.__version__.split(".")[:2]))
+            self.return_type = "tensor" if v >= (4, 54) else "tuple"
+        assert self.return_type is not None
+        super().__init__(qcfg, fp_name=fp_name)
+
+        # Child QuantConfigs -------------------------------------------------
+        attn_cfg = qcfg.child("self_attn") if qcfg else None
+        mlp_cfg = qcfg.child("mlp") if qcfg else None
+
+        # Quantized sub-modules ---------------------------------------------
+        assert hasattr(fp_layer, "self_attn") and isinstance(
+            fp_layer.self_attn, torch.nn.Module
+        )
+        assert hasattr(fp_layer, "mlp") and isinstance(fp_layer.mlp, torch.nn.Module)
+        self.self_attn = PTQWrapper(
+            fp_layer.self_attn, qcfg=attn_cfg, fp_name=f"{fp_name}.self_attn"
+        )
+        self.mlp = PTQWrapper(fp_layer.mlp, qcfg=mlp_cfg, fp_name=f"{fp_name}.mlp")
+
+        # LayerNorms remain FP (copied from fp_layer to keep weights)
+        assert hasattr(fp_layer, "input_layernorm") and isinstance(
+            fp_layer.input_layernorm, torch.nn.Module
+        )
+        assert hasattr(fp_layer, "post_attention_layernorm") and isinstance(
+            fp_layer.post_attention_layernorm, torch.nn.Module
+        )
+        self.input_layernorm = fp_layer.input_layernorm
+        self.post_attention_layernorm = fp_layer.post_attention_layernorm
+
+        # Static causal mask template ---------------------------------------
+        assert hasattr(fp_layer.self_attn, "config") and hasattr(
+            fp_layer.self_attn.config, "max_position_embeddings"
+        )
+        assert isinstance(fp_layer.self_attn.config.max_position_embeddings, int)
+        max_seq = fp_layer.self_attn.config.max_position_embeddings
+        mask = torch.full((1, 1, max_seq, max_seq), float("-120"))
+        mask.triu_(1)
+        self.register_buffer("causal_mask_template", mask, persistent=False)
+
+    def _slice_causal(self, seq_len: int, device: torch.device) -> torch.Tensor:
+        """Return `[1,1,L,L]` causal mask slice on *device*."""
+        assert isinstance(self.causal_mask_template, torch.Tensor)
+        return self.causal_mask_template[..., :seq_len, :seq_len].to(device)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional["Cache"] = None,  # type: ignore[name-defined]
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor] | torch.Tensor:
+        if output_attentions:
+            raise NotImplementedError(
+                "QuantLlamaDecoderLayer does not support output attention yet."
+            )
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+
+        if attention_mask is None or attention_mask.dtype == torch.bool:
+            L = hidden_states.size(1)
+            attention_mask = self._slice_causal(L, hidden_states.device)
+
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # ─── MLP block ─────────────────────────────────────────────────
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        if self.return_type == "tuple":
+            return (hidden_states,)
+        elif self.return_type == "tensor":
+            return hidden_states
+        else:
+            raise RuntimeError("Invalid return type.")
+
+    # No local observers; just recurse into children
+    def _all_observers(self):
+        yield from self.self_attn._all_observers()
+        yield from self.mlp._all_observers()

tico/experimental/quantization/ptq/wrappers/registry.py

@@ -29,6 +29,7 @@ _CORE_MODULES = (
     "tico.experimental.quantization.ptq.wrappers.nn.quant_linear",
     "tico.experimental.quantization.ptq.wrappers.nn.quant_silu",
     "tico.experimental.quantization.ptq.wrappers.llama.quant_attn",
+    "tico.experimental.quantization.ptq.wrappers.llama.quant_decoder_layer",
     "tico.experimental.quantization.ptq.wrappers.llama.quant_mlp",
     # add future core wrappers here
 )

tico/passes/remove_redundant_assert_nodes.py

@@ -21,7 +21,9 @@ from tico.utils.utils import is_target_node
 
 
 assert_node_targets = [
+    torch.ops.aten._assert_scalar.default,
     torch.ops.aten._assert_tensor_metadata.default,
+    torch.ops.aten.sym_constrain_range_for_size.default,  # Related to symbolic shape validation
 ]
 
 
@@ -29,7 +31,7 @@ assert_node_targets = [
 class RemoveRedundantAssertionNodes(PassBase):
     """
     This removes redundant assertion nodes.
-    - `aten.assert_tensor_meta.default`
+    When assertion node is erased, related comparison nodes are also removed by graph.eliminate_dead_code().
     """
 
     def __init__(self):

{tico-0.1.0.dev250831.dist-info → tico-0.1.0.dev250902.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250831
+Version: 0.1.0.dev250902
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250831.dist-info → tico-0.1.0.dev250902.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=AGSSd8oWeJ_KCyfANsnPaH1Sf2i33iYNqoeT153KsC0,1883
+tico/__init__.py,sha256=PXZzhb0ZexNIwGhVJpg4Ln_RqskbSIMigqj0GdZgbeA,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
@@ -62,8 +62,10 @@ tico/experimental/quantization/ptq/mode.py,sha256=lT-T8vIv8YWcwrjT7xXVhOw1g7aoAd
 tico/experimental/quantization/ptq/qscheme.py,sha256=uwhv7bCxOOXB3I-IKlRyr_u4eXOq48uIqGy4TLDqGxY,1301
 tico/experimental/quantization/ptq/quant_config.py,sha256=nm7570Y1X2mOT_8s27ilWid04otor6cVTi9GwgAEaKc,4300
 tico/experimental/quantization/ptq/examples/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
+tico/experimental/quantization/ptq/examples/compare_ppl.py,sha256=ODaRB234iy2dFfGIBd-OtKxdSzxnIbgKZkQ_o30tSts,5287
 tico/experimental/quantization/ptq/examples/quantize_linear.py,sha256=8zq-ZJDYgam0xQ-PbC6Xb1I7W1mv0Wi-b--IP2wwXtw,4539
 tico/experimental/quantization/ptq/examples/quantize_llama_attn.py,sha256=cVWUSSzaZWFp5QZkNkrlpHU3kXyP84QtnZbahVml_yQ,4329
+tico/experimental/quantization/ptq/examples/quantize_llama_decoder_layer.py,sha256=mBWrjkyEovYQsPC4Rrsri6Pm1rlFmDb3NiP0DQQhFyM,5751
 tico/experimental/quantization/ptq/examples/quantize_llama_mlp.py,sha256=N1qZQgt1S-xZrdv-PW7OfXEcv0gsO2q9faOF4aD-zKo,4147
 tico/experimental/quantization/ptq/observers/__init__.py,sha256=WF2MvL9M_jl-B1FqcY9zic34NOCRp17HkRYv-TMxMr4,613
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
@@ -72,15 +74,17 @@ tico/experimental/quantization/ptq/observers/ema.py,sha256=MAMdBmjVNMg_vsqXrcBzb
 tico/experimental/quantization/ptq/observers/identity.py,sha256=vkec8Or-7VwM4zkFEvEKROQJk8XEHMVX8mBNDnxSyS8,2591
 tico/experimental/quantization/ptq/observers/minmax.py,sha256=mLHkwIzWFzQXev7EU7w1333KckwRjukc3_cUPJOnUfs,1486
 tico/experimental/quantization/ptq/observers/mx.py,sha256=aP4qmBgeiRIYZJksShN5gs6UyYOFi2-Sbk5k5xvPQ4w,1863
-tico/experimental/quantization/ptq/utils/__init__.py,sha256=PL9IZgiWoMtsXVljeOy7KymmLVP238SXEFRLXYK72WQ,126
+tico/experimental/quantization/ptq/utils/__init__.py,sha256=MrQwMbbKS0dJrO8jsceCai4Z59iKQNpTPZND3GN6TrM,216
+tico/experimental/quantization/ptq/utils/metrics.py,sha256=EW_FQmJrl9Y4esspZQ0GHfJ58RwuJUz0l8IfYq3NWY4,4461
 tico/experimental/quantization/ptq/utils/reduce_utils.py,sha256=3kWawLB91EcvvHlCrNqqfZF7tpgr22htBSA049mKw_4,973
 tico/experimental/quantization/ptq/wrappers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 tico/experimental/quantization/ptq/wrappers/ptq_wrapper.py,sha256=F9sK_DiRaXiGNHULcwIbs5EUtHz6ZJ7N4r5CWTTfhsM,2442
 tico/experimental/quantization/ptq/wrappers/quant_elementwise.py,sha256=LhEoobfvto6zKrBOKL4gmxfFFc31jHzyQV_zfps-iQM,3604
 tico/experimental/quantization/ptq/wrappers/quant_module_base.py,sha256=vkcDos_knGSS29rIZuEIWkAJLHrENbGz8nCH2-iara8,5969
-tico/experimental/quantization/ptq/wrappers/registry.py,sha256=TwH-MD-qkTkG6M-f1VqFLmSNcXLNYsh21yjyzCcojJc,4706
-tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=b360gkQ0RxExmiV-ZaaxwJdMPJ53g6uCRlR2-_dOby0,240
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=M1D_foC0PR-Ii4G0lbOO3_pmhvHlMF28NolK_q2DZtw,4783
+tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=4xuAYnJcohMTtBzrH4cxq8WKG2GQo8nbhektVg8w7F0,380
 tico/experimental/quantization/ptq/wrappers/llama/quant_attn.py,sha256=WIUI6EFMTvvruvqu8pBxWy6qJeDyjkaYbJk1R3pAmwE,8578
+tico/experimental/quantization/ptq/wrappers/llama/quant_decoder_layer.py,sha256=2XsIf5rcabDXXkahqriSxfo2curFq0Y5bnRPcYkJPg8,7187
 tico/experimental/quantization/ptq/wrappers/llama/quant_mlp.py,sha256=uZMnrX66oZwxhKhcNbLXXeri-WxxRBiZnr15aBXJMm0,3562
 tico/experimental/quantization/ptq/wrappers/nn/__init__.py,sha256=I9uTt5HfcRoMEDYHpAeATMv2TbCQiX0ZbfUFMzSJ4Qw,336
 tico/experimental/quantization/ptq/wrappers/nn/quant_layernorm.py,sha256=G5Sgt-tXnzh0Rxyk-2honmZIfEQOZlRfOsoDBdSGmA4,6887
@@ -117,7 +121,7 @@ tico/passes/lower_to_slice.py,sha256=OzlFzK3lBYyYwC3WThsWd94Ob4JINIJF8UaLAtnumzU
 tico/passes/merge_consecutive_cat.py,sha256=ayZNLDA1DFM7Fxxi2Dmk1CujkgUuaVCH1rhQgLrvvOQ,2701
 tico/passes/ops.py,sha256=cSj3Sk2x2cOE9b8oU5pmSa_rHr-iX2lORzu3N_UHMSQ,2967
 tico/passes/remove_nop.py,sha256=Hf91p_EJAOC6DyWNthash0_UWtEcNc_M7znamQfYQ5Y,2686
-tico/passes/remove_redundant_assert_nodes.py,sha256=IONd3xBy6I8tH6_Y1eN3_eCHH7WTC8soBgjXzOju9cQ,1612
+tico/passes/remove_redundant_assert_nodes.py,sha256=rYbTCyuNIXIC-2NreHKBVCuaSUkEQvB_iSRzb26P_EA,1821
 tico/passes/remove_redundant_expand.py,sha256=auyqIoQT4HJhiJfuUe6BrEtUhvz221ohnIK5EuszWeg,2112
 tico/passes/remove_redundant_permute.py,sha256=98UsaZzFZdQzEEAR1pIzRisAf6hgfXLa88aayjalt3E,4292
 tico/passes/remove_redundant_reshape.py,sha256=aeep6LDvY58GEuOrWckkEXnJa6wkkbiJ9FrimT9F3-s,16384
@@ -240,9 +244,9 @@ 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.dev250831.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250831.dist-info/METADATA,sha256=1tkuUVzG_pExk9oBSJO3s5opMoRAVMkWns4biwGPDu0,8450
-tico-0.1.0.dev250831.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250831.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250831.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250831.dist-info/RECORD,,
+tico-0.1.0.dev250902.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250902.dist-info/METADATA,sha256=CePT5yw5-ln0-Ct8n61iGDnFfnoASlqAfPQmxRQ9QQ0,8450
+tico-0.1.0.dev250902.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250902.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250902.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250902.dist-info/RECORD,,