tico 0.1.0.dev250904__py3-none-any.whl → 0.1.0.dev251109__py3-none-any.whl

tico/quantization/wrapq/examples/debug_quant_outputs.py

@@ -0,0 +1,224 @@
+# 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.
+
+# ============================================================================
+# 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.
+# ============================================================================
+
+import argparse
+import sys
+
+import torch
+import tqdm
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from tico.quantization import convert, prepare
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.wrapq.utils.introspection import (
+    build_fqn_map,
+    compare_layer_outputs,
+    save_fp_outputs,
+)
+from tico.quantization.wrapq.wrappers.ptq_wrapper import PTQWrapper
+
+# 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,
+}
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+    "float16": torch.float16,
+}
+
+# Hardcoded dataset settings
+DATASET_NAME = "wikitext"
+DATASET_CONFIG = "wikitext-2-raw-v1"
+TRAIN_SPLIT = "train"
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Layer-wise diff debugging pipeline for PTQ"
+    )
+    parser.add_argument(
+        "--model", type=str, required=True, help="HF repo name or local path."
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda" if torch.cuda.is_available() else "cpu",
+        help="Device to run on (cuda|cpu|mps).",
+    )
+    parser.add_argument(
+        "--dtype",
+        choices=list(DTYPE_MAP.keys()),
+        default="float32",
+        help=f"Model dtype for load.",
+    )
+    parser.add_argument(
+        "--stride",
+        type=int,
+        default=512,
+        help="Sliding-window stride used during calibration.",
+    )
+    parser.add_argument(
+        "--calib-preset",
+        choices=list(TOKENS.keys()),
+        default="debug",
+        help="Calibration token budget preset.",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="Random seed.")
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Enable only if you trust the model repo code.",
+    )
+    parser.add_argument(
+        "--hf-token",
+        type=str,
+        default=None,
+        help="Optional HF token for gated/private repos.",
+    )
+    parser.add_argument(
+        "--use-cache",
+        dest="use_cache",
+        action="store_true",
+        default=False,
+        help="Use model KV cache if enabled (off by default).",
+    )
+    parser.add_argument(
+        "--no-tqdm", action="store_true", help="Disable tqdm progress bars."
+    )
+
+    args = parser.parse_args()
+
+    # Basic setup
+    torch.manual_seed(args.seed)
+    device = torch.device(args.device)
+    dtype = DTYPE_MAP[args.dtype]  # noqa: E999 (kept readable)
+
+    print("=== Config ===")
+    print(f"Model            : {args.model}")
+    print(f"Device           : {device.type}")
+    print(f"DType            : {args.dtype}")
+    print(f"Stride           : {args.stride}")
+    print(
+        f"Calib preset     : {args.calib_preset} ({TOKENS[args.calib_preset]:,} tokens)"
+    )
+    print(f"Use HF cache?    : {args.use_cache}")
+    print()
+
+    # -------------------------------------------------------------------------
+    # 1. Load the FP backbone and tokenizer
+    # -------------------------------------------------------------------------
+    print("Loading FP model …")
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model,
+        trust_remote_code=args.trust_remote_code,
+        token=args.hf_token,
+    )
+    model = (
+        AutoModelForCausalLM.from_pretrained(
+            args.model,
+            torch_dtype=dtype,
+            trust_remote_code=args.trust_remote_code,
+            token=args.hf_token,
+        )
+        .to(device)
+        .eval()
+    )
+
+    # Disable KV cache to force full forward passes for introspection
+    model.config.use_cache = args.use_cache
+
+    # Build module -> FQN map before wrapping
+    m_to_fqn = build_fqn_map(model)
+
+    # Prepare calibration inputs (HF Wikitext-2 train split)
+    CALIB_TOKENS = TOKENS[args.calib_preset]
+    print(f"Calibrating with {CALIB_TOKENS:,} tokens.\n")
+    # Use Wikitext-2 train split for calibration.
+    dataset = load_dataset(DATASET_NAME, DATASET_CONFIG, split=TRAIN_SPLIT)
+
+    # -------------------------------------------------------------------------
+    # 2. Wrap every layer with PTQWrapper (UINT-8 activations)
+    # -------------------------------------------------------------------------
+    print("Wrapping layers with PTQWrapper …")
+    qcfg = PTQConfig()  # default: per-tensor UINT8
+    prepare(model, qcfg)
+
+    # -------------------------------------------------------------------------
+    # 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)
+
+    # Save reference FP activations before observers clamp/quantize
+    save_handles, act_cache = save_fp_outputs(model)
+
+    iterator = range(0, ids.size(1) - 1, args.stride)
+    if not args.no_tqdm:
+        iterator = tqdm.tqdm(iterator, desc="Act-Calibration")
+    with torch.no_grad():
+        for i in iterator:
+            inputs = ids[:, i : i + args.stride]
+            model(inputs)  # observers collect act. ranges
+
+    # Remove save hooks now that FP activations are cached
+    for h in save_handles:
+        h.remove()
+
+    # Freeze (scale, zero-point) after calibration
+    convert(model)
+
+    # Register diff hooks and measure per-layer deltas
+    cmp_handles = compare_layer_outputs(model, act_cache, metrics=["diff", "peir"])
+    # Use same inputs for comparison.
+    with torch.no_grad():
+        model(inputs)
+
+    assert isinstance(cmp_handles, list)
+    for h in cmp_handles:
+        h.remove()
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as e:
+        print(f"\n[Error] {e}", file=sys.stderr)
+        sys.exit(1)

tico/{experimental/quantization/ptq → quantization/wrapq}/examples/quantize_linear.py

@@ -29,13 +29,15 @@ import pathlib
 import torch
 import torch.nn as nn
 
-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.nn.quant_linear import QuantLinear
+from tico.quantization import convert, prepare
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.evaluation.metric import compute_peir
+from tico.quantization.evaluation.utils import plot_two_outputs
+from tico.quantization.wrapq.mode import Mode
+from tico.quantization.wrapq.wrappers.nn.quant_linear import QuantLinear
 from tico.utils.utils import SuppressWarning
 
+
 # -------------------------------------------------------------------------
 # 0. Define a toy model (1 Linear layer only)
 # -------------------------------------------------------------------------
@@ -60,20 +62,19 @@ fp32_layer = model.fc
 # -------------------------------------------------------------------------
 # 1. Replace the Linear with QuantLinear wrapper
 # -------------------------------------------------------------------------
-model.fc = QuantLinear(fp32_layer)  # type: ignore[assignment]
-# model.fc = PTQWrapper(fp32_layer) (Wrapping helper class)
+model.fc = prepare(fp32_layer, PTQConfig())  # type: ignore[assignment]
 qlayer = model.fc  # alias for brevity
 
 # -------------------------------------------------------------------------
 # 2. Single-pass calibration (collect activation ranges)
 # -------------------------------------------------------------------------
-assert isinstance(qlayer, QuantLinear)
+assert isinstance(qlayer.wrapped, QuantLinear)
 with torch.no_grad():
-    qlayer.enable_calibration()
     for _ in range(16):  # small toy batch
         x = torch.randn(4, 16)  # (batch=4, features=16)
         _ = model(x)
-    qlayer.freeze_qparams()  # lock scales & zero-points
+
+convert(qlayer)
 
 assert qlayer._mode is Mode.QUANT, "Quantization mode should be active now."
 

tico/{experimental/quantization/ptq → quantization/wrapq}/examples/quantize_llama_attn.py

@@ -17,13 +17,12 @@ 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_attn import (
-    QuantLlamaAttention,
-)
+from tico.quantization import convert, prepare
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.evaluation.metric import compute_peir
+from tico.quantization.evaluation.utils import plot_two_outputs
+from tico.quantization.wrapq.mode import Mode
+from tico.quantization.wrapq.wrappers.llama.quant_attn import QuantLlamaAttention
 from tico.utils.utils import SuppressWarning
 
 name = "Maykeye/TinyLLama-v0"
@@ -34,12 +33,11 @@ tokenizer = AutoTokenizer.from_pretrained(name)
 # 1. Replace layer-0’s MLP with QuantLlamaMLP
 # -------------------------------------------------------------------------
 orig_attn = model.model.layers[0].self_attn
-model.model.layers[0].self_attn = QuantLlamaAttention(
-    orig_attn
-)  # PTQWrapper(orig_attn) is also fine
+model.model.layers[0].self_attn = prepare(orig_attn, PTQConfig())
 model.eval()
 
 attn_q = model.model.layers[0].self_attn  # quant wrapper
+assert isinstance(attn_q.wrapped, QuantLlamaAttention)
 rotary = model.model.rotary_emb
 
 # -------------------------------------------------------------------------
@@ -55,7 +53,6 @@ PROMPTS = [
 ]
 
 with torch.no_grad():
-    attn_q.enable_calibration()
     for prompt in PROMPTS:
         ids = tokenizer(prompt, return_tensors="pt")
         embeds = model.model.embed_tokens(ids["input_ids"])
@@ -63,7 +60,8 @@ with torch.no_grad():
         S = cos_sin[0].shape[1]
         float_mask = torch.zeros(1, 1, S, S)
         _ = attn_q(embeds, cos_sin)  # observers collect
-    attn_q.freeze_qparams()
+
+convert(attn_q)
 
 assert attn_q._mode is Mode.QUANT, "Quantization mode should be active now."
 

tico/{experimental/quantization/ptq → quantization/wrapq}/examples/quantize_llama_decoder_layer.py

@@ -31,10 +31,12 @@ 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 (
+from tico.quantization import convert, prepare
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.evaluation.metric import compute_peir
+from tico.quantization.evaluation.utils import plot_two_outputs
+from tico.quantization.wrapq.mode import Mode
+from tico.quantization.wrapq.wrappers.llama.quant_decoder_layer import (
     QuantLlamaDecoderLayer,
 )
 from tico.utils.utils import SuppressWarning
@@ -50,12 +52,11 @@ 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.model.layers[0] = prepare(fp32_layer, PTQConfig())
 model.eval()
 
 qlayer = model.model.layers[0]  # alias for brevity
+assert isinstance(qlayer.wrapped, QuantLlamaDecoderLayer)
 
 # -------------------------------------------------------------------------
 # 2. Single-pass calibration (gather activation ranges)
@@ -70,7 +71,6 @@ PROMPTS = [
 ]
 
 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"])
@@ -78,7 +78,8 @@ with torch.no_grad():
         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()
+
+convert(qlayer)
 
 assert qlayer._mode is Mode.QUANT, "Quantization mode should be active now."
 

tico/{experimental/quantization/ptq → quantization/wrapq}/examples/quantize_llama_mlp.py

@@ -18,13 +18,14 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 import tico
-from tico.experimental.quantization.evaluation.metric import compute_peir
-from tico.experimental.quantization.evaluation.utils import plot_two_outputs
-from tico.experimental.quantization.ptq.dtypes import INT16
-from tico.experimental.quantization.ptq.mode import Mode
-from tico.experimental.quantization.ptq.qscheme import QScheme
-from tico.experimental.quantization.ptq.quant_config import QuantConfig
-from tico.experimental.quantization.ptq.wrappers.llama.quant_mlp import QuantLlamaMLP
+from tico.quantization import convert, prepare
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.evaluation.metric import compute_peir
+from tico.quantization.evaluation.utils import plot_two_outputs
+from tico.quantization.wrapq.dtypes import INT16
+from tico.quantization.wrapq.mode import Mode
+from tico.quantization.wrapq.qscheme import QScheme
+from tico.quantization.wrapq.wrappers.llama.quant_mlp import QuantLlamaMLP
 from tico.utils.utils import SuppressWarning
 
 name = "Maykeye/TinyLLama-v0"
@@ -36,13 +37,13 @@ model.eval()
 # 1. Replace layer-0’s MLP with QuantLlamaMLP
 # -------------------------------------------------------------------------
 fp32_mlp = model.model.layers[0].mlp
-model.model.layers[0].mlp = QuantLlamaMLP(
-    fp32_mlp,
-    qcfg=QuantConfig(default_dtype=INT16, default_qscheme=QScheme.PER_TENSOR_SYMM),
-)  # PTQWrapper(fp32_mlp) is also fine
+model.model.layers[0].mlp = prepare(
+    fp32_mlp, PTQConfig(default_dtype=INT16, default_qscheme=QScheme.PER_TENSOR_SYMM)
+)
 model.eval()
 
 mlp_q = model.model.layers[0].mlp
+assert isinstance(mlp_q.wrapped, QuantLlamaMLP)
 
 # -------------------------------------------------------------------------
 # 2. Single-pass calibration
@@ -57,13 +58,12 @@ PROMPTS = [
 ]
 
 with torch.no_grad():
-    mlp_q.enable_calibration()
     for prompt in PROMPTS:
         enc = tokenizer(prompt, return_tensors="pt")
         emb = model.model.embed_tokens(enc["input_ids"])
         _ = mlp_q(emb)
 
-    mlp_q.freeze_qparams()
+convert(mlp_q)
 
 assert mlp_q._mode is Mode.QUANT, "Quantization mode should be active now."
 

tico/quantization/wrapq/examples/quantize_with_gptq.py

@@ -0,0 +1,265 @@
+# 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.
+# =============================================================================
+
+import argparse
+import sys
+from typing import Any
+
+import torch
+import tqdm
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from tico.quantization import convert, prepare
+from tico.quantization.config.gptq import GPTQConfig
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.wrapq.observers.affine_base import AffineObserverBase
+from tico.quantization.wrapq.utils.introspection import build_fqn_map
+from tico.quantization.wrapq.utils.metrics import perplexity
+from tico.quantization.wrapq.wrappers.ptq_wrapper import PTQWrapper
+from tico.quantization.wrapq.wrappers.quant_module_base import QuantModuleBase
+
+
+# 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,
+}
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+    "float16": torch.float16,
+}
+
+# Hardcoded dataset settings
+DATASET_NAME = "wikitext"
+DATASET_CONFIG = "wikitext-2-raw-v1"
+TRAIN_SPLIT = "train"
+TEST_SPLIT = "test"
+
+# -------------------------------------------------------------------------
+# 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)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="GPTQ+PTQ pipeline (weight-only + activation UINT8)"
+    )
+    parser.add_argument(
+        "--model", type=str, required=True, help="HF repo name or local path."
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda" if torch.cuda.is_available() else "cpu",
+        help="Device to run on (cuda|cpu|mps).",
+    )
+    parser.add_argument(
+        "--dtype",
+        choices=list(DTYPE_MAP.keys()),
+        default="float32",
+        help="Model dtype for load.",
+    )
+    parser.add_argument(
+        "--stride",
+        type=int,
+        default=512,
+        help="Sliding-window stride used for calibration and eval.",
+    )
+    parser.add_argument(
+        "--calib-preset",
+        choices=list(TOKENS.keys()),
+        default="debug",
+        help="Activation calibration token budget preset.",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="Random seed.")
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Enable only if you trust the model repo code.",
+    )
+    parser.add_argument(
+        "--hf-token",
+        type=str,
+        default=None,
+        help="Optional HF token for gated/private repos.",
+    )
+    parser.add_argument(
+        "--use-cache",
+        dest="use_cache",
+        action="store_true",
+        default=False,
+        help="Use model KV cache if enabled (off by default).",
+    )
+    parser.add_argument(
+        "--no-tqdm", action="store_true", help="Disable tqdm progress bars."
+    )
+
+    args = parser.parse_args()
+
+    # Basic setup
+    torch.manual_seed(args.seed)
+    device = torch.device(args.device)
+    dtype = DTYPE_MAP[args.dtype]
+
+    print("=== Config ===")
+    print(f"Model            : {args.model}")
+    print(f"Device           : {device.type}")
+    print(f"DType            : {args.dtype}")
+    print(f"Stride           : {args.stride}")
+    print(
+        f"Calib preset     : {args.calib_preset} ({TOKENS[args.calib_preset]:,} tokens)"
+    )
+    print(f"Use HF cache?    : {args.use_cache}")
+    print()
+
+    # -------------------------------------------------------------------------
+    # 2. Load the FP backbone and tokenizer
+    # -------------------------------------------------------------------------
+    print("Loading FP model …")
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model,
+        trust_remote_code=args.trust_remote_code,
+        token=args.hf_token,
+    )
+    model = (
+        AutoModelForCausalLM.from_pretrained(
+            args.model,
+            torch_dtype=dtype,
+            trust_remote_code=args.trust_remote_code,
+            token=args.hf_token,
+        )
+        .to(device)
+        .eval()
+    )
+
+    model.config.use_cache = args.use_cache
+
+    # Build module -> FQN map BEFORE wrapping
+    m_to_fqn = build_fqn_map(model)
+
+    # -------------------------------------------------------------------------
+    # 3. Run GPTQ (weight-only) pass
+    # -------------------------------------------------------------------------
+    print("Applying GPTQ …")
+    dataset_test = load_dataset(DATASET_NAME, DATASET_CONFIG, split=TEST_SPLIT)
+    q_m = prepare(model, GPTQConfig(), inplace=True)
+
+    it = (
+        dataset_test
+        if args.no_tqdm
+        else tqdm.tqdm(dataset_test, desc="GPTQ calibration")
+    )
+    for d in it:
+        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)
+    # -------------------------------------------------------------------------
+    print("Wrapping layers with PTQWrapper …")
+    qcfg = PTQConfig()  # default: per-tensor UINT8
+    prepare(q_m, qcfg)
+
+    # -------------------------------------------------------------------------
+    # 5. Single-pass activation calibration
+    # -------------------------------------------------------------------------
+    print("Calibrating UINT-8 observers …")
+    CALIB_TOKENS = TOKENS[args.calib_preset]
+    print(f"Calibrating with {CALIB_TOKENS:,} tokens.\n")
+    dataset_train = load_dataset(DATASET_NAME, DATASET_CONFIG, split=TRAIN_SPLIT)
+    calib_txt = " ".join(dataset_train["text"])[:CALIB_TOKENS]
+    train_ids = tokenizer(calib_txt, return_tensors="pt").input_ids.to(device)
+
+    # Overwrite weight observers with GPTQ statistics
+    if hasattr(q_m, "quantizers") and isinstance(q_m.quantizers, dict):
+        inject_gptq_qparams(q_m, q_m.quantizers)
+    else:
+        print(
+            "[Warn] q_m.quantizers not found or not a dict; skipping GPTQ qparam injection."
+        )
+
+    # Forward passes to collect activation ranges
+    iterator = range(0, train_ids.size(1) - 1, args.stride)
+    if not args.no_tqdm:
+        iterator = tqdm.tqdm(iterator, desc="Act-calibration")
+    with torch.no_grad():
+        for i in iterator:
+            q_m(train_ids[:, i : i + args.stride])
+
+    # Freeze all Q-params (scale, zero-point)
+    convert(q_m)
+
+    # -------------------------------------------------------------------------
+    # 6. Evaluate perplexity on Wikitext-2
+    # -------------------------------------------------------------------------
+    print("\nCalculating perplexities …")
+    enc = tokenizer("\n\n".join(dataset_test["text"]), return_tensors="pt")
+    ppl_uint8 = perplexity(q_m, enc, device, stride=args.stride)
+
+    print("\n┌── Wikitext-2 test perplexity ─────────────")
+    print(f"│ UINT-8 : {ppl_uint8:8.2f}")
+    print("└───────────────────────────────────────────")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as e:
+        print(f"\n[Error] {e}", file=sys.stderr)
+        sys.exit(1)