tico 0.1.0.dev250915__py3-none-any.whl → 0.1.0.dev250917__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250915"
+__version__ = "0.1.0.dev250917"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

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

@@ -77,7 +77,7 @@ def main():
         "--dtype",
         choices=list(DTYPE_MAP.keys()),
         default="float32",
-        help="Model dtype for load (float32|bfloat16|float16).",
+        help=f"Model dtype for load.",
     )
     parser.add_argument(
         "--stride", type=int, default=512, help="Sliding-window stride for perplexity."
@@ -126,7 +126,9 @@ def main():
     print(f"DType            : {args.dtype}")
     print(f"Stride           : {args.stride}")
     print(f"Use HF cache?    : {args.use_cache}")
-    print(f"Calib preset     : {args.calib_preset}")
+    print(
+        f"Calib preset     : {args.calib_preset} ({TOKENS[args.calib_preset]:,} tokens)"
+    )
     print()
 
     # -------------------------------------------------------------------------

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

@@ -12,19 +12,6 @@
 # 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
 # ----------------------------------------------------------------------------
@@ -43,12 +30,21 @@ from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
 # 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
+import argparse
+import sys
+
+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
 
 # Token-budget presets for activation calibration
 TOKENS: dict[str, int] = {
@@ -59,71 +55,185 @@ TOKENS: dict[str, int] = {
     # 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)
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+    "float16": torch.float16,
+}
 
-# (a) Enable CALIB mode on every QuantModuleBase
-for l in model.model.layers:
-    l.enable_calibration()
+# Hardcoded dataset settings
+DATASET_NAME = "wikitext"
+DATASET_CONFIG = "wikitext-2-raw-v1"
+TRAIN_SPLIT = "train"
 
-# 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
+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."
+    )
 
-# Remove save hooks now that FP activations are cached
-for h in save_handles:
-    h.remove()
+    args = parser.parse_args()
 
-# (b) Freeze (scale, zero-point) after calibration
-for l in model.model.layers:
-    l.freeze_qparams()
+    # Basic setup
+    torch.manual_seed(args.seed)
+    device = torch.device(args.device)
+    dtype = DTYPE_MAP[args.dtype]  # noqa: E999 (kept readable)
 
-# (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)
+    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()
+    )
 
-assert isinstance(cmp_handles, list)
-for h in cmp_handles:
-    h.remove()
+    # 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 = 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)
+
+    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()
+
+    # (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.
+    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/utils/signature.py

@@ -141,22 +141,21 @@ class ModelInputSpec:
         args = flatten_and_convert_args(args)
         kwargs = flatten_and_convert_kwargs(kwargs)
 
+        arg_num = len(args) + len(kwargs)
+        m_input_num = len(self.names)
+        if arg_num != m_input_num:
+            raise ValueError(
+                f"Mismatch: number of model inputs and number of passed arguments are not the same: inputs({m_input_num}) != passed({arg_num}), input spec: {self.names}"
+            )
+
         # 1. positional arguments
         for i, val in enumerate(args):
-            if i >= len(self.names):
-                raise ValueError(f"Too many positional arguments ({i+1}).")
             name = self.names[i]
-            if name in kwargs:
-                raise TypeError(
-                    f"Got multiple values for argument '{name}' (positional and keyword)."
-                )
             inputs.append(val)
 
         # 2. keyword arguments
         for idx in range(len(args), len(self.names)):
             name = self.names[idx]
-            if name not in kwargs:
-                raise ValueError(f"Missing argument for input '{name}'.")
             inputs.append(kwargs[name])
 
         if check:

{tico-0.1.0.dev250915.dist-info → tico-0.1.0.dev250917.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250915
+Version: 0.1.0.dev250917
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250915.dist-info → tico-0.1.0.dev250917.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=RT4YNN5EM4rbOVOWo1BHEO8vnfPWLEwVNjMFh3qRYeY,1883
+tico/__init__.py,sha256=Da7Ln6MuWCBJXrjts6OsAslWSS79toVgPG2PITYPzE0,1883
 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 +62,8 @@ 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=QWUuO50lITnooYqEe57VV6mvIHKWZMB_TOGvtZ8C8qQ,8238
-tico/experimental/quantization/ptq/examples/debug_quant_outputs.py,sha256=astXzx-maq1W4gKvX2QaGmD2Tpmjunv4JqDYVk9eZRQ,5177
+tico/experimental/quantization/ptq/examples/compare_ppl.py,sha256=eVQn8-M24QkoCy_FCBQLSlUrnyqUDSkvUFpUpEdpMx4,8265
+tico/experimental/quantization/ptq/examples/debug_quant_outputs.py,sha256=Hpx_jj46WISwdVp33NrIadizVAzf1nFTXuAcHsKEQuQ,8179
 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
@@ -243,7 +243,7 @@ tico/utils/pytree_utils.py,sha256=jrk3N6X6LiUnBCX_gM1K9nywbVAJBVnszlTAgeIeDUc,52
 tico/utils/record_input.py,sha256=QN-8D71G_WAX3QQQ5CIwbEfFJZTQ3CvL4wCMiVddua4,3894
 tico/utils/register_custom_op.py,sha256=895SKZeXQzolK-mPG38cQC37Be76xUV_Ujw1k1ts9_w,28218
 tico/utils/serialize.py,sha256=mEuusEzi82WFsz3AkowgWwxSLeo50JDxyOj6yYDQhEI,1914
-tico/utils/signature.py,sha256=R2GV0alRpXEbZISqPKyxCUWbgDcsrQ2ovbVG3737IzA,9595
+tico/utils/signature.py,sha256=3OOwyVJzfcGcgC0LiVmOcUIzfqSk27qoNHhkoCI7zPY,9530
 tico/utils/torch_compat.py,sha256=oc6PztVsXdHcQ3iaVR90wLLxrGaj6zFHWZ8K9rRS6q8,1795
 tico/utils/trace_decorators.py,sha256=ddLIiKQfSaQrxgF1kNpwjFTQnXENzeSfcr1kuAW4jGI,3221
 tico/utils/utils.py,sha256=aySftYnNTsqVAMcGs_3uX3-hz577a2cj4p1aVV-1XeQ,12747
@@ -252,9 +252,9 @@ 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.dev250915.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250915.dist-info/METADATA,sha256=5l-EgJKZwF179QnVqWApKdARhQxw0c2iibtckWUu-XA,8450
-tico-0.1.0.dev250915.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250915.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250915.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250915.dist-info/RECORD,,
+tico-0.1.0.dev250917.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250917.dist-info/METADATA,sha256=WJdcwQ8suuOhdWCv9cW8_RW_qyckaOM5jEzlvi00vbM,8450
+tico-0.1.0.dev250917.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250917.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250917.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250917.dist-info/RECORD,,