tico 0.1.0.dev250803__py3-none-any.whl → 0.1.0.dev251102__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/quantization/wrapq/examples/quantize_linear.py

@@ -0,0 +1,107 @@
+# 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 — Simple Linear Model
+# -----------------------------------------------------------------------------
+# This demo shows a minimal PTQ flow for a toy model:
+#   1. Define a simple model with a single Linear layer.
+#   2. Replace the FP32 Linear with a QuantLinear wrapper.
+#   3. Run a short calibration pass to collect activation statistics.
+#   4. Freeze scales / zero-points and switch to INT-simulation mode.
+#   5. Compare INT vs FP32 outputs with a mean-absolute-diff check.
+#   6. Export the quantized model to a Circle format.
+# =============================================================================
+
+import pathlib
+
+import torch
+import torch.nn as nn
+
+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)
+# -------------------------------------------------------------------------
+class TinyLinearModel(nn.Module):
+    """A minimal model: single Linear layer."""
+
+    def __init__(self, in_features=16, out_features=8):
+        super().__init__()
+        self.fc = nn.Linear(in_features, out_features, bias=False)
+
+    def forward(self, x):
+        return self.fc(x)
+
+
+# Instantiate FP32 model
+model = TinyLinearModel()
+model.eval()
+
+# Keep FP32 reference for diff check
+fp32_layer = model.fc
+
+# -------------------------------------------------------------------------
+# 1. Replace the Linear with QuantLinear wrapper
+# -------------------------------------------------------------------------
+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.wrapped, QuantLinear)
+with torch.no_grad():
+    for _ in range(16):  # small toy batch
+        x = torch.randn(4, 16)  # (batch=4, features=16)
+        _ = model(x)
+
+convert(qlayer)
+
+assert qlayer._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick INT-sim vs FP32 sanity check
+# -------------------------------------------------------------------------
+x = torch.randn(2, 16)
+with torch.no_grad():
+    int8_out = model(x)
+    fp32_out = fp32_layer(x)
+
+print("┌───────────── Quantization Error Summary ─────────────")
+print(f"│ Mean |diff|: {(int8_out - fp32_out).abs().mean().item():.6f}")
+print(f"│ PEIR       : {compute_peir(fp32_out, int8_out) * 100:.6f} %")
+print("└──────────────────────────────────────────────────────")
+print(plot_two_outputs(fp32_out, int8_out))
+
+# -------------------------------------------------------------------------
+# 4. Export the calibrated model to Circle
+# -------------------------------------------------------------------------
+import tico
+
+save_path = pathlib.Path("tiny_linear.q.circle")
+example_input = torch.randn(1, 16)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(model, (example_input,))  # forward(x) only
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")

tico/quantization/wrapq/examples/quantize_llama_attn.py

@@ -0,0 +1,101 @@
+# 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 pathlib
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+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"
+model = AutoModelForCausalLM.from_pretrained(name)
+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 = 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
+
+# -------------------------------------------------------------------------
+# 2. Single-pass calibration
+# -------------------------------------------------------------------------
+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():
+    for prompt in PROMPTS:
+        ids = tokenizer(prompt, return_tensors="pt")
+        embeds = model.model.embed_tokens(ids["input_ids"])
+        cos_sin = rotary(embeds, ids["input_ids"])
+        S = cos_sin[0].shape[1]
+        float_mask = torch.zeros(1, 1, S, S)
+        _ = attn_q(embeds, cos_sin)  # observers collect
+
+convert(attn_q)
+
+assert attn_q._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick diff check (INT-sim vs FP32)
+# -------------------------------------------------------------------------
+ids = tokenizer("check", return_tensors="pt")
+emb = model.model.embed_tokens(ids["input_ids"])
+pos = rotary(emb, ids["input_ids"])
+S = pos[0].shape[1]
+float_mask = torch.zeros(1, 1, S, S)
+with torch.no_grad():
+    int8 = attn_q(emb, pos)[0]
+    fp32 = orig_attn(emb, position_embeddings=pos, attention_mask=None)[0]
+
+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 quantized block
+# -------------------------------------------------------------------------
+import tico
+
+save_path = pathlib.Path("attn.q.circle")
+B, S, D = 1, 4, model.config.hidden_size
+example = torch.randn(B, S, D)
+example_pos = rotary(example, torch.arange(S)[None, :])
+float_mask = torch.zeros(1, 1, S, S)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(attn_q, (example, example_pos))
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")

tico/quantization/wrapq/examples/quantize_llama_decoder_layer.py

@@ -0,0 +1,125 @@
+# 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.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
+
+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] = 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)
+# -------------------------------------------------------------------------
+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():
+    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)
+
+convert(qlayer)
+
+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/quantization/wrapq/examples/quantize_llama_mlp.py

@@ -0,0 +1,95 @@
+# 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 pathlib
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+import tico
+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"
+model = AutoModelForCausalLM.from_pretrained(name)
+tokenizer = AutoTokenizer.from_pretrained(name)
+model.eval()
+
+# -------------------------------------------------------------------------
+# 1. Replace layer-0’s MLP with QuantLlamaMLP
+# -------------------------------------------------------------------------
+fp32_mlp = model.model.layers[0].mlp
+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
+# -------------------------------------------------------------------------
+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():
+    for prompt in PROMPTS:
+        enc = tokenizer(prompt, return_tensors="pt")
+        emb = model.model.embed_tokens(enc["input_ids"])
+        _ = mlp_q(emb)
+
+convert(mlp_q)
+
+assert mlp_q._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick diff check (INT-sim vs FP32)
+# -------------------------------------------------------------------------
+with torch.no_grad():
+    ids = tokenizer("quant all tensors!", return_tensors="pt")
+    emb = model.model.embed_tokens(ids["input_ids"])
+    int16 = mlp_q(emb)  # INT-sim
+    fp32 = fp32_mlp(emb)  # baseline reference
+
+print("┌───────────── Quantization Error Summary ─────────────")
+print(f"│ Mean |diff|: {(int16 - fp32).abs().mean().item():.6f}")
+print(f"│ PEIR       : {compute_peir(fp32, int16) * 100:.6f} %")
+print("└──────────────────────────────────────────────────────")
+print(plot_two_outputs(fp32, int16))
+
+# -------------------------------------------------------------------------
+# 4. Export the quantized block
+# -------------------------------------------------------------------------
+save_path = pathlib.Path("mlp.q.circle")
+example_in = (torch.randn(1, 1, model.config.hidden_size),)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(mlp_q, example_in)
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")