tico 0.1.0.dev250902__py3-none-any.whl → 0.1.0.dev250904__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250902"
+__version__ = "0.1.0.dev250904"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/ptq/__init__.py

@@ -1,13 +1 @@
-"""
-Public PTQ API — re-export the most common symbols.
-"""
-
-from tico.experimental.quantization.ptq.dtypes import DType
-from tico.experimental.quantization.ptq.mode import Mode
-from tico.experimental.quantization.ptq.qscheme import QScheme
-
-__all__ = [
-    "DType",
-    "Mode",
-    "QScheme",
-]
+# DO NOT REMOVE THIS FILE

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

@@ -26,7 +26,7 @@ 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.utils.metrics import perplexity
 from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
 
 # -------------------------------------------------------------------------

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

@@ -0,0 +1,129 @@
+# 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.
+
+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.introspection import (
+    build_fqn_map,
+    compare_layer_outputs,
+    save_fp_outputs,
+)
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+
+# ============================================================================
+# LAYER-WISE DIFF DEBUGGING PIPELINE
+# ----------------------------------------------------------------------------
+# A quantization debugging pipeline that identifies accuracy regressions
+# by comparing UINT vs FP outputs at each layer.
+#
+#   1. Load a full-precision (FP) LLaMA-3-1B model.
+#   2. Wrap each Transformer block with PTQWrapper (activations → fake-quant).
+#   3. Capture reference FP layer outputs before quantization.
+#   4. Calibrate UINT-8 activation observers in a single pass.
+#   5. Freeze quantization parameters (scale, zero-point).
+#   6. Re-run inference and compare UINT-8 vs FP outputs per layer.
+#   7. Report where quantization hurts the most.
+#
+# Use this pipeline to trace precision loss layer by layer, and pinpoint
+# problematic modules during post-training quantization.
+# ============================================================================
+
+# -------------------------------------------------------------------------
+# 0. Global configuration
+# -------------------------------------------------------------------------
+MODEL_NAME = "meta-llama/Meta-Llama-3-1B"
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+STRIDE = 512
+
+# 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 the FP backbone
+# -------------------------------------------------------------------------
+print("Loading FP model …")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME).to(DEVICE).eval()
+model.config.use_cache = False  # disable KV-cache → full forward
+m_to_fqn = build_fqn_map(model)  # map modules → fully-qualified names
+
+# Use Wikitext-2 train split for calibration.
+dataset = load_dataset("wikitext", "wikitext-2-raw-v1", split="train")
+
+# -------------------------------------------------------------------------
+# 2. Wrap every layer with PTQWrapper (UINT-8 activations)
+# -------------------------------------------------------------------------
+print("Wrapping layers with PTQWrapper …")
+qcfg = QuantConfig()  # default: per-tensor UINT8
+
+new_layers = torch.nn.ModuleList()
+for idx, fp_layer in enumerate(model.model.layers):
+    layer_cfg = qcfg.child(f"layer{idx}")
+    q_layer = PTQWrapper(
+        fp_layer,
+        qcfg=layer_cfg,
+        fp_name=m_to_fqn.get(fp_layer),
+    )
+    new_layers.append(q_layer)
+
+model.model.layers = new_layers  # swap in quant wrappers
+
+# -------------------------------------------------------------------------
+# 3. Activation calibration plus FP-vs-UINT8 diffing
+# -------------------------------------------------------------------------
+print("Calibrating UINT-8 observers …")
+calib_txt = " ".join(dataset["text"])[:CALIB_TOKENS]
+ids = tokenizer(calib_txt, return_tensors="pt").input_ids.to(DEVICE)
+
+# (a) Enable CALIB mode on every QuantModuleBase
+for l in model.model.layers:
+    l.enable_calibration()
+
+# Save reference FP activations before observers clamp/quantize
+save_handles, act_cache = save_fp_outputs(model)
+
+with torch.no_grad():
+    for i in tqdm.trange(0, ids.size(1) - 1, STRIDE, desc="Act-calibration"):
+        inputs = ids[:, i : i + STRIDE]
+        model(inputs)  # observers collect act. ranges
+
+# Remove save hooks now that FP activations are cached
+for h in save_handles:
+    h.remove()
+
+# (b) Freeze (scale, zero-point) after calibration
+for l in model.model.layers:
+    l.freeze_qparams()
+
+# (c) Register diff hooks and measure per-layer deltas
+cmp_handles = compare_layer_outputs(model, act_cache, metrics=["diff", "peir"])
+# Use same inputs for comparison.
+model(inputs)
+
+assert isinstance(cmp_handles, list)
+for h in cmp_handles:
+    h.remove()

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

@@ -0,0 +1,165 @@
+# 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.
+
+# =============================================================================
+# PTQ + GPTQ HYBRID QUANTIZATION PIPELINE
+# -----------------------------------------------------------------------------
+# This script shows how to:
+#   1. Load a pretrained FP Llama-3 model.
+#   2. Run GPTQ to quantize weights only.
+#   3. Wrap every Transformer layer with a PTQWrapper to quantize activations.
+#   4. Calibrate UINT-8 observers in a single pass over a text corpus.
+#   5. Inject GPTQ’s per-tensor weight scales / zero-points into the PTQ graph.
+#   6. Freeze all Q-params and compute Wikitext-2 perplexity.
+# =============================================================================
+
+from typing import Any
+
+import torch
+import tqdm
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from tico.experimental.quantization import convert, prepare
+from tico.experimental.quantization.config import GPTQConfig
+from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.utils.introspection import build_fqn_map
+from tico.experimental.quantization.ptq.utils.metrics import perplexity
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+
+# -------------------------------------------------------------------------
+# 0. Global configuration
+# -------------------------------------------------------------------------
+MODEL_NAME = "meta-llama/Meta-Llama-3-1B"
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+STRIDE = 512
+
+# 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"]
+
+# -------------------------------------------------------------------------
+# 1. Helper — copy GPTQ (scale, zp) into PTQ observers
+# -------------------------------------------------------------------------
+def inject_gptq_qparams(
+    root: torch.nn.Module,
+    gptq_quantizers: dict[str, Any],  # {fp_name: quantizer}
+    weight_obs_name: str = "weight",
+):
+    """
+    For every `QuantModuleBase` whose `fp_name` matches a GPTQ key,
+    locate the observer called `weight_obs_name` and overwrite its
+    (scale, zero-point), then lock them against further updates.
+    """
+    for m in root.modules():
+        if not isinstance(m, QuantModuleBase):
+            continue
+        if m.fp_name is None:
+            continue
+        quantizer = gptq_quantizers.get(m.fp_name)
+        if quantizer is None:
+            continue
+        obs = m.get_observer(weight_obs_name)
+        if obs is None:
+            continue
+        assert isinstance(obs, AffineObserverBase)
+        # GPTQ quantizer attributes
+        obs.load_qparams(quantizer.scale, quantizer.zero, lock=True)
+
+
+# -------------------------------------------------------------------------
+# 2. Load the FP backbone
+# -------------------------------------------------------------------------
+print("Loading FP model …")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME).to(DEVICE).eval()
+model.config.use_cache = False  # disable KV-cache → full forward
+m_to_fqn = build_fqn_map(model)  # map modules → fully-qualified names
+
+# -------------------------------------------------------------------------
+# 3. Run GPTQ (weight-only) pass
+# -------------------------------------------------------------------------
+print("Applying GPTQ …")
+dataset = load_dataset("wikiText", "wikitext-2-raw-v1", split="test")
+q_m = prepare(model, GPTQConfig(), inplace=True)
+
+for d in tqdm.tqdm(dataset, desc="GPTQ calibration"):
+    ids = tokenizer(d["text"], return_tensors="pt").input_ids.to(DEVICE)
+    q_m(ids)  # observers gather weight stats
+
+q_m = convert(q_m, inplace=True)  # materialize INT-weight tensors
+
+# -------------------------------------------------------------------------
+# 4. Wrap every layer with PTQWrapper (activation UINT-8)
+# -------------------------------------------------------------------------
+qcfg = QuantConfig()  # default: per-tensor UINT8
+new_layers = torch.nn.ModuleList()
+
+for idx, fp_layer in enumerate(q_m.model.layers):
+    layer_cfg = qcfg.child(f"layer{idx}")
+    q_layer = PTQWrapper(
+        fp_layer,
+        qcfg=layer_cfg,
+        fp_name=m_to_fqn.get(fp_layer),
+    )
+    new_layers.append(q_layer)
+
+q_m.model.layers = new_layers
+
+# -------------------------------------------------------------------------
+# 5. 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) Enable CALIB mode on every QuantModuleBase
+for l in q_m.model.layers:
+    l.enable_calibration()
+
+# (b) Overwrite weight observers with GPTQ statistics
+inject_gptq_qparams(q_m, q_m.quantizers)
+
+with torch.no_grad():
+    for i in tqdm.trange(0, ids.size(1) - 1, STRIDE, desc="Act-calibration"):
+        q_m(ids[:, i : i + STRIDE])  # observers collect act. ranges
+
+# (c) Freeze all Q-params (scale, zp)
+for l in q_m.model.layers:
+    l.freeze_qparams()
+
+# -------------------------------------------------------------------------
+# 6. 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")
+ppl_uint8 = perplexity(q_m, enc, DEVICE, stride=STRIDE)
+
+print("\n┌── Wikitext-2 test perplexity ─────────────")
+print(f"│ UINT-8 : {ppl_uint8:8.2f}")
+print("└───────────────────────────────────────────")

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

@@ -1,15 +1 @@
-from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
-from tico.experimental.quantization.ptq.observers.base import ObserverBase
-from tico.experimental.quantization.ptq.observers.ema import EMAObserver
-from tico.experimental.quantization.ptq.observers.identity import IdentityObserver
-from tico.experimental.quantization.ptq.observers.minmax import MinMaxObserver
-from tico.experimental.quantization.ptq.observers.mx import MXObserver
-
-__all__ = [
-    "AffineObserverBase",
-    "ObserverBase",
-    "EMAObserver",
-    "IdentityObserver",
-    "MinMaxObserver",
-    "MXObserver",
-]
+# DO NOT REMOVE THIS FILE

tico/experimental/quantization/ptq/observers/ema.py

@@ -15,7 +15,7 @@
 import torch
 
 from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
-from tico.experimental.quantization.ptq.utils import channelwise_minmax
+from tico.experimental.quantization.ptq.utils.reduce_utils import channelwise_minmax
 
 
 class EMAObserver(AffineObserverBase):

tico/experimental/quantization/ptq/observers/minmax.py

@@ -15,7 +15,7 @@
 import torch
 
 from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
-from tico.experimental.quantization.ptq.utils import channelwise_minmax
+from tico.experimental.quantization.ptq.utils.reduce_utils import channelwise_minmax
 
 
 class MinMaxObserver(AffineObserverBase):

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

@@ -1,7 +1 @@
-from tico.experimental.quantization.ptq.utils.metrics import perplexity
-from tico.experimental.quantization.ptq.utils.reduce_utils import channelwise_minmax
-
-__all__ = [
-    "channelwise_minmax",
-    "perplexity",
-]
+# DO NOT REMOVE THIS FILE

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

@@ -0,0 +1,169 @@
+# 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 Callable, Dict, List, Optional, Tuple
+
+import torch
+
+from tico.experimental.quantization.evaluation.metric import MetricCalculator
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+
+
+def build_fqn_map(root: torch.nn.Module) -> dict[torch.nn.Module, str]:
+    """
+    Return {module_object: full_qualified_name} without touching the modules.
+    """
+    return {m: n for n, m in root.named_modules()}
+
+
+def save_fp_outputs(
+    model: torch.nn.Module,
+) -> Tuple[List[torch.utils.hooks.RemovableHandle], Dict[str, torch.Tensor]]:
+    """
+    Register forward-hooks on every `QuantModuleBase` wrapper itself (not the
+    wrapped `module`) and cache its output while the wrapper runs in CALIB mode.
+
+    Parameters
+    ----------
+    model : torch.nn.Module
+        The model whose wrappers are already switched to CALIB mode
+        (`enable_calibration()` has been called).
+
+    Returns
+    -------
+    handles : list[RemovableHandle]
+        Hook handles; call `.remove()` on each one to detach the hooks.
+    cache : dict[str, torch.Tensor]
+        Mapping "wrapper-name → cached FP32 activation" captured from the first
+        forward pass. Keys default to `wrapper.fp_name`; if that attribute is
+        `None`, the `id(wrapper)` string is used instead.
+    """
+    cache: Dict[str, torch.Tensor] = {}
+    handles: List[torch.utils.hooks.RemovableHandle] = []
+
+    def _save(name: str):
+        def hook(_, __, out: torch.Tensor | Tuple):
+            if isinstance(out, tuple):
+                out = out[0]
+            assert isinstance(out, torch.Tensor)
+            cache[name] = out.detach()
+
+        return hook
+
+    for m in model.modules():
+        if isinstance(m, QuantModuleBase):
+            name = m.fp_name or str(id(m))
+            handles.append(m.register_forward_hook(_save(name)))
+
+    return handles, cache
+
+
+def compare_layer_outputs(
+    model: torch.nn.Module,
+    cache: Dict[str, torch.Tensor],
+    *,
+    metrics: Optional[List[str]] = None,
+    custom_metrics: Optional[Dict[str, Callable]] = None,
+    rtol: float = 1e-3,
+    atol: float = 1e-3,
+    collect: bool = False,
+):
+    """
+    Register forward-hooks on every `QuantModuleBase` wrapper to compare its
+    QUANT-mode output to the FP32 reference saved by `save_fp_outputs()`.
+
+    Each hook prints a per-layer diff report:
+
+        ✓  layer_name  max=1.23e-02  mean=8.45e-04     (within tolerance)
+        ⚠️ layer_name  max=3.07e+00  mean=5.12e-01     (exceeds tolerance)
+
+    Parameters
+    ----------
+    model : torch.nn.Module
+        The model whose wrappers are now in QUANT mode
+        (`freeze_qparams()` has been called).
+    cache : dict[str, torch.Tensor]
+        The reference activations captured during CALIB mode.
+    metrics
+        Metrics to compute. Defaults to `["diff"]`. Add `peir` to print PEIR.
+    custom_metrics
+        Optional user metric functions. Same signature as built-ins.
+    rtol, atol : float, optional
+        Relative / absolute tolerances used to flag large deviations
+        (similar to `torch.allclose` semantics).
+    collect : bool, optional
+        • False (default) → print one-line report per layer, return `None`
+        • True            → suppress printing, return a nested dict
+                                {layer_name -> {metric -> value}}
+
+    Returns
+    -------
+    handles
+        Hook handles; call `.remove()` once diffing is complete.
+    results
+        Only if *collect* is True.
+    """
+    metrics = metrics or ["diff"]
+    calc = MetricCalculator(custom_metrics)
+    handles: List[torch.utils.hooks.RemovableHandle] = []
+    results: Dict[
+        str, Dict[str, float]
+    ] = {}  # Dict[layer_name, Dict[metric_name, value]]
+
+    def _cmp(name: str):
+        ref = cache.get(name)
+
+        def hook(_, __, out):
+            if ref is None:
+                if not collect:
+                    print(f"[{name}]  no cached reference")
+                return
+            if isinstance(out, tuple):
+                out = out[0]
+            assert isinstance(out, torch.Tensor)
+
+            # Compute all requested metrics
+            res = calc.compute([ref], [out], metrics)  # lists with length-1 tensors
+            res = {k: v[0] for k, v in res.items()}  # flatten
+
+            if collect:
+                results[name] = res  # type: ignore[assignment]
+                return
+
+            # Pretty print ------------------------------------------------ #
+            diff_val = res.get("diff") or res.get("max_abs_diff")
+            thresh = atol + rtol * ref.abs().max().item()
+            flag = "⚠️" if (diff_val is not None and diff_val > thresh) else "✓"  # type: ignore[operator]
+
+            pieces = [f"{flag} {name:45s}"]
+            for key, val in res.items():
+                pieces.append(f"{key}={val:<7.4}")
+            print("  ".join(pieces))
+
+        return hook
+
+    for m in model.modules():
+        if isinstance(m, PTQWrapper):
+            # skip the internal fp module inside the wrapper
+            continue
+        if isinstance(m, QuantModuleBase):
+            lname = m.fp_name or str(id(m))
+            handles.append(m.register_forward_hook(_cmp(lname)))
+
+    if collect:
+        return handles, results
+    return handles

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

@@ -0,0 +1 @@
+# DO NOT REMOVE THIS FILE

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

@@ -1,9 +1 @@
-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", "QuantLlamaDecoderLayer", "QuantLlamaMLP"]
+# DO NOT REMOVE THIS FILE

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

@@ -25,7 +25,10 @@ from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
 from tico.experimental.quantization.ptq.wrappers.registry import try_register
 
 
-@try_register("transformers.models.llama.modeling_llama.LlamaAttention")
+@try_register(
+    "transformers.models.llama.modeling_llama.LlamaAttention",
+    "transformers.models.llama.modeling_llama.LlamaSdpaAttention",
+)
 class QuantLlamaAttention(QuantModuleBase):
     def __init__(
         self,

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

@@ -1,11 +1 @@
-from tico.experimental.quantization.ptq.wrappers.nn.quant_layernorm import (
-    QuantLayerNorm,
-)
-from tico.experimental.quantization.ptq.wrappers.nn.quant_linear import QuantLinear
-from tico.experimental.quantization.ptq.wrappers.nn.quant_silu import QuantSiLU
-
-__all__ = [
-    "QuantLayerNorm",
-    "QuantLinear",
-    "QuantSiLU",
-]
+# DO NOT REMOVE THIS FILE

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

@@ -90,7 +90,9 @@ def register(
 
 
 # ───────────────────────────── conditional decorator
-def try_register(path: str) -> Callable[[Type[QuantModuleBase]], Type[QuantModuleBase]]:
+def try_register(
+    *paths: str,
+) -> Callable[[Type[QuantModuleBase]], Type[QuantModuleBase]]:
     """
     @try_register("transformers.models.llama.modeling_llama.LlamaMLP")
 
@@ -99,14 +101,15 @@ def try_register(path: str) -> Callable[[Type[QuantModuleBase]], Type[QuantModul
     """
 
     def _decorator(quant_cls: Type[QuantModuleBase]):
-        module_name, _, cls_name = path.rpartition(".")
-        try:
-            mod = importlib.import_module(module_name)
-            fp_cls = getattr(mod, cls_name)
-            _WRAPPERS[fp_cls] = quant_cls
-        except (ModuleNotFoundError, AttributeError):
-            # transformers not installed or class renamed – silently skip
-            pass
+        for path in paths:
+            module_name, _, cls_name = path.rpartition(".")
+            try:
+                mod = importlib.import_module(module_name)
+                fp_cls = getattr(mod, cls_name)
+                _WRAPPERS[fp_cls] = quant_cls
+            except (ModuleNotFoundError, AttributeError):
+                # optional dep missing or class renamed – skip silently
+                pass
         return quant_cls
 
     return _decorator

{tico-0.1.0.dev250902.dist-info → tico-0.1.0.dev250904.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250902
+Version: 0.1.0.dev250904
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250902.dist-info → tico-0.1.0.dev250904.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=PXZzhb0ZexNIwGhVJpg4Ln_RqskbSIMigqj0GdZgbeA,1883
+tico/__init__.py,sha256=Q7A_sZgX9kbJSpo2ndJB19BlbFSGnajl9HAFGt8D2Q0,1883
 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
@@ -56,37 +56,40 @@ tico/experimental/quantization/passes/propagate_qparam_backward.py,sha256=TGtyW0
 tico/experimental/quantization/passes/propagate_qparam_forward.py,sha256=RhUHGCR2RpBO5KYkQ7Z8U5u7HEwDq2wdKHLKAJCi-5c,5138
 tico/experimental/quantization/passes/quantize_bias.py,sha256=T7YxJ70N0tSK0FF9VJZA5iP0sHdnnsX9GX4AT4JDFSk,4325
 tico/experimental/quantization/passes/remove_weight_dequant_op.py,sha256=gI1MtrHazWpdNfys7f1ngTTWplzluF7SA-uX0HMR5Mc,6592
-tico/experimental/quantization/ptq/__init__.py,sha256=ZoPdEwZ1i1n5pBFChx8GuUrkfRP2vsSoLPNILQjNBaA,298
+tico/experimental/quantization/ptq/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/ptq/dtypes.py,sha256=xfCBtq6mQmUYRwsoFgII6gvRl1raQi0Inj9pznDuKwQ,2236
 tico/experimental/quantization/ptq/mode.py,sha256=lT-T8vIv8YWcwrjT7xXVhOw1g7aoAdh_3PWB-ptPKaI,1052
 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/compare_ppl.py,sha256=SmSmaCBVWTcGRPRk2zopDqESD_gF8D7J4kUNNZ-0cMk,5295
+tico/experimental/quantization/ptq/examples/debug_quant_outputs.py,sha256=astXzx-maq1W4gKvX2QaGmD2Tpmjunv4JqDYVk9eZRQ,5177
 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/examples/quantize_with_gptq.py,sha256=w21Qao5_6SnWMuxmnZbZOoqaLQOuSnK52mHin4aedtA,6979
+tico/experimental/quantization/ptq/observers/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
 tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
-tico/experimental/quantization/ptq/observers/ema.py,sha256=MAMdBmjVNMg_vsqXrcBzbw_1nFJ-j4Gz651k3-VlaMQ,2057
+tico/experimental/quantization/ptq/observers/ema.py,sha256=oISP1XaD3lapVaHQKscD3rjLcKbhOy4Nvi6dqRFZwF8,2070
 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/minmax.py,sha256=WWcAyEIrd5j3k9qsoBJi3nUnWtrwPaKlR9CPezbDSqQ,1499
 tico/experimental/quantization/ptq/observers/mx.py,sha256=aP4qmBgeiRIYZJksShN5gs6UyYOFi2-Sbk5k5xvPQ4w,1863
-tico/experimental/quantization/ptq/utils/__init__.py,sha256=MrQwMbbKS0dJrO8jsceCai4Z59iKQNpTPZND3GN6TrM,216
+tico/experimental/quantization/ptq/utils/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
+tico/experimental/quantization/ptq/utils/introspection.py,sha256=y2oGf7RoApMHJeXLmIz3VVWB9vazGEgyLbxLiVTTQdw,6000
 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/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 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=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/registry.py,sha256=wauoZdZBR15bGj1Upt9owEfFDT-Tj6HzciG9HDM1BHo,4845
+tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
+tico/experimental/quantization/ptq/wrappers/llama/quant_attn.py,sha256=-K1COLHIHfJZhQu-RE6KfJIkaL7S6yR4iUj48QkjMTw,8652
 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/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/ptq/wrappers/nn/quant_layernorm.py,sha256=G5Sgt-tXnzh0Rxyk-2honmZIfEQOZlRfOsoDBdSGmA4,6887
 tico/experimental/quantization/ptq/wrappers/nn/quant_linear.py,sha256=xW-VEPB7RJoslS3xLVCdhIuMjppknvpkZleRGK4JFVQ,2240
 tico/experimental/quantization/ptq/wrappers/nn/quant_silu.py,sha256=XnJDggkWUTfXC1-BLeAbcCUtp687XLIkIIbuQlqycDw,1864
@@ -244,9 +247,9 @@ 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.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,,
+tico-0.1.0.dev250904.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250904.dist-info/METADATA,sha256=jCLzVEpVnwflcRRdt3AZElB_Lw1_lWSzusXsiA5RAig,8450
+tico-0.1.0.dev250904.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250904.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250904.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250904.dist-info/RECORD,,