tico 0.1.0.dev250826__py3-none-any.whl → 0.1.0.dev250828__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250826"
+__version__ = "0.1.0.dev250828"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

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

@@ -0,0 +1,103 @@
+# 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.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.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 = QuantLlamaAttention(
+    orig_attn
+)  # PTQWrapper(orig_attn) is also fine
+model.eval()
+
+attn_q = model.model.layers[0].self_attn  # quant wrapper
+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():
+    attn_q.enable_calibration()
+    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
+    attn_q.freeze_qparams()
+
+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/experimental/quantization/ptq/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.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.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 = QuantLlamaMLP(
+    fp32_mlp,
+    qcfg=QuantConfig(default_dtype=INT16, default_qscheme=QScheme.PER_TENSOR_SYMM),
+)  # PTQWrapper(fp32_mlp) is also fine
+model.eval()
+
+mlp_q = model.model.layers[0].mlp
+
+# -------------------------------------------------------------------------
+# 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():
+    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()
+
+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()}")

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

@@ -0,0 +1,6 @@
+from tico.experimental.quantization.ptq.wrappers.llama.quant_attn import (
+    QuantLlamaAttention,
+)
+from tico.experimental.quantization.ptq.wrappers.llama.quant_mlp import QuantLlamaMLP
+
+__all__ = ["QuantLlamaAttention", "QuantLlamaMLP"]

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

@@ -0,0 +1,236 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import try_register
+
+
+@try_register("transformers.models.llama.modeling_llama.LlamaAttention")
+class QuantLlamaAttention(QuantModuleBase):
+    def __init__(
+        self,
+        fp_attn: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+
+        cfg = fp_attn.config
+        assert hasattr(cfg, "hidden_size") and hasattr(cfg, "num_attention_heads")
+        assert hasattr(cfg, "num_key_value_heads")
+        assert isinstance(cfg.hidden_size, int) and isinstance(
+            cfg.num_attention_heads, int
+        )
+        assert isinstance(cfg.num_key_value_heads, int)
+        self.hdim = getattr(cfg, "head_dim", cfg.hidden_size // cfg.num_attention_heads)
+        self.kv_rep = cfg.num_attention_heads // cfg.num_key_value_heads
+
+        # constant scale (1/√d)
+        self.scale_t = torch.tensor(self.hdim**-0.5)
+        self.obs_scale = self._make_obs("scale")
+
+        # ---- wrap q k v o projections via PTQWrapper ---------------
+        q_cfg = qcfg.child("q_proj") if qcfg else None
+        k_cfg = qcfg.child("k_proj") if qcfg else None
+        v_cfg = qcfg.child("v_proj") if qcfg else None
+        o_cfg = qcfg.child("o_proj") if qcfg else None
+        assert hasattr(fp_attn, "q_proj") and isinstance(
+            fp_attn.q_proj, torch.nn.Module
+        )
+        assert hasattr(fp_attn, "k_proj") and isinstance(
+            fp_attn.k_proj, torch.nn.Module
+        )
+        assert hasattr(fp_attn, "v_proj") and isinstance(
+            fp_attn.v_proj, torch.nn.Module
+        )
+        assert hasattr(fp_attn, "o_proj") and isinstance(
+            fp_attn.o_proj, torch.nn.Module
+        )
+        self.q_proj = PTQWrapper(
+            fp_attn.q_proj, qcfg=q_cfg, fp_name=f"{fp_name}.q_proj"
+        )
+        self.k_proj = PTQWrapper(
+            fp_attn.k_proj, qcfg=k_cfg, fp_name=f"{fp_name}.k_proj"
+        )
+        self.v_proj = PTQWrapper(
+            fp_attn.v_proj, qcfg=v_cfg, fp_name=f"{fp_name}.v_proj"
+        )
+        self.o_proj = PTQWrapper(
+            fp_attn.o_proj, qcfg=o_cfg, fp_name=f"{fp_name}.o_proj"
+        )
+
+        # ---- create arithmetic observers ---------------------------
+        mk = self._make_obs
+        self.obs_hidden = mk("hidden")
+
+        self.obs_cos = mk("cos")
+        self.obs_sin = mk("sin")
+
+        self.obs_causal_mask = mk("causal_mask")
+
+        # rotate-half sub-steps
+        self.obs_q_x1 = mk("q_x1")
+        self.obs_q_x2 = mk("q_x2")
+        self.obs_q_neg = mk("q_neg")
+        self.obs_q_cat = mk("q_cat")
+        self.obs_k_x1 = mk("k_x1")
+        self.obs_k_x2 = mk("k_x2")
+        self.obs_k_neg = mk("k_neg")
+        self.obs_k_cat = mk("k_cat")
+
+        # q / k paths
+        self.obs_q_cos = mk("q_cos")
+        self.obs_q_sin = mk("q_sin")
+        self.obs_q_rot = mk("q_rot")
+        self.obs_k_cos = mk("k_cos")
+        self.obs_k_sin = mk("k_sin")
+        self.obs_k_rot = mk("k_rot")
+
+        # logits / softmax / out
+        self.obs_logits_raw = mk("logits_raw")
+        self.obs_logits = mk("logits")
+        self.obs_mask_add = mk("mask_add")
+        self.obs_softmax = mk("softmax")
+        self.obs_attn_out = mk("attn_out")
+
+        # Static causal mask template
+        assert hasattr(cfg, "max_position_embeddings")
+        max_seq = cfg.max_position_embeddings
+        mask = torch.full((1, 1, max_seq, max_seq), float("-120"))  # type: ignore[arg-type]
+        mask.triu_(1)
+        self.register_buffer("causal_mask_template", mask, persistent=False)
+
+    def _rot(self, t, o_x1, o_x2, o_neg, o_cat):
+        x1, x2 = torch.chunk(t, 2, dim=-1)
+        x1 = self._fq(x1, o_x1)
+        x2 = self._fq(x2, o_x2)
+        x2n = self._fq(-x2, o_neg)
+        return self._fq(torch.cat((x2n, x1), -1), o_cat)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value=None,  # not supported yet
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        if past_key_value is not None:
+            raise NotImplementedError(
+                "QuantLlamaAttention does not support KV cache yet."
+            )
+
+        hidden = self._fq(hidden_states, self.obs_hidden)
+        B, S, _ = hidden.shape
+        H = self.hdim
+
+        # projections
+        q = self.q_proj(hidden).view(B, S, -1, H).transpose(1, 2)
+        k = self.k_proj(hidden).view(B, S, -1, H).transpose(1, 2)
+        v = self.v_proj(hidden).view(B, S, -1, H).transpose(1, 2)
+
+        # rope tables
+        cos, sin = position_embeddings
+        cos = self._fq(cos, self.obs_cos)
+        sin = self._fq(sin, self.obs_sin)
+        cos_u, sin_u = cos.unsqueeze(1), sin.unsqueeze(1)
+
+        # q_rot
+        q_half = self._rot(
+            q, self.obs_q_x1, self.obs_q_x2, self.obs_q_neg, self.obs_q_cat
+        )
+        q_cos = self._fq(q * cos_u, self.obs_q_cos)
+        q_sin = self._fq(q_half * sin_u, self.obs_q_sin)
+        q_rot = self._fq(q_cos + q_sin, self.obs_q_rot)
+
+        # k_rot
+        k_half = self._rot(
+            k, self.obs_k_x1, self.obs_k_x2, self.obs_k_neg, self.obs_k_cat
+        )
+        k_cos = self._fq(k * cos_u, self.obs_k_cos)
+        k_sin = self._fq(k_half * sin_u, self.obs_k_sin)
+        k_rot = self._fq(k_cos + k_sin, self.obs_k_rot)
+
+        # logits
+        k_rep = k_rot.repeat_interleave(self.kv_rep, dim=1)
+        logits_raw = self._fq(q_rot @ k_rep.transpose(-2, -1), self.obs_logits_raw)
+        scale = self._fq(self.scale_t, self.obs_scale)
+        logits = self._fq(logits_raw * scale, self.obs_logits)
+
+        if attention_mask is None or attention_mask.dtype == torch.bool:
+            _, _, q_len, k_len = logits.shape
+            assert isinstance(self.causal_mask_template, torch.Tensor)
+            attention_mask = self.causal_mask_template[..., :q_len, :k_len].to(
+                hidden_states.device
+            )
+        attention_mask = self._fq(attention_mask, self.obs_causal_mask)
+        logits = self._fq(logits + attention_mask, self.obs_mask_add)
+
+        # softmax
+        attn_weights = torch.softmax(logits, -1, dtype=torch.float32).to(q.dtype)
+        attn_weights = self._fq(attn_weights, self.obs_softmax)
+
+        # attn out
+        v_rep = v.repeat_interleave(self.kv_rep, dim=1)
+        attn_out = (
+            self._fq(attn_weights @ v_rep, self.obs_attn_out)
+            .transpose(1, 2)
+            .reshape(B, S, -1)
+        )
+
+        # final projection
+        return self.o_proj(attn_out), attn_weights
+
+    def _all_observers(self):
+        # local first
+        yield from (
+            self.obs_hidden,
+            self.obs_scale,
+            self.obs_cos,
+            self.obs_sin,
+            self.obs_causal_mask,
+            self.obs_q_x1,
+            self.obs_q_x2,
+            self.obs_q_neg,
+            self.obs_q_cat,
+            self.obs_k_x1,
+            self.obs_k_x2,
+            self.obs_k_neg,
+            self.obs_k_cat,
+            self.obs_q_cos,
+            self.obs_q_sin,
+            self.obs_q_rot,
+            self.obs_k_cos,
+            self.obs_k_sin,
+            self.obs_k_rot,
+            self.obs_logits_raw,
+            self.obs_logits,
+            self.obs_mask_add,
+            self.obs_softmax,
+            self.obs_attn_out,
+        )
+        # recurse into children that are QuantModuleBase
+        for m in (self.q_proj, self.k_proj, self.v_proj, self.o_proj):
+            yield from m._all_observers()

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

@@ -0,0 +1,98 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import try_register
+
+
+@try_register("transformers.models.llama.modeling_llama.LlamaMLP")
+class QuantLlamaMLP(QuantModuleBase):
+    def __init__(
+        self,
+        mlp_fp: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+
+        # ----- child configs (hierarchical override) -------------------
+        gate_cfg = qcfg.child("gate_proj") if qcfg else None
+        up_cfg = qcfg.child("up_proj") if qcfg else None
+        down_cfg = qcfg.child("down_proj") if qcfg else None
+        act_cfg = qcfg.child("act_fn") if qcfg else None
+
+        # ----- wrap three Linear layers -------------------------------
+        assert hasattr(mlp_fp, "gate_proj") and isinstance(
+            mlp_fp.gate_proj, torch.nn.Module
+        )
+        assert hasattr(mlp_fp, "up_proj") and isinstance(
+            mlp_fp.up_proj, torch.nn.Module
+        )
+        assert hasattr(mlp_fp, "down_proj") and isinstance(
+            mlp_fp.down_proj, torch.nn.Module
+        )
+        self.gate_proj = PTQWrapper(
+            mlp_fp.gate_proj, qcfg=gate_cfg, fp_name=f"{fp_name}.gate_proj"
+        )
+        self.up_proj = PTQWrapper(
+            mlp_fp.up_proj, qcfg=up_cfg, fp_name=f"{fp_name}.up_proj"
+        )
+        self.down_proj = PTQWrapper(
+            mlp_fp.down_proj, qcfg=down_cfg, fp_name=f"{fp_name}.down_proj"
+        )
+
+        # ----- activation ---------------------------------------------
+        assert hasattr(mlp_fp, "act_fn") and isinstance(mlp_fp.act_fn, torch.nn.Module)
+        self.act_fn = PTQWrapper(
+            mlp_fp.act_fn, qcfg=act_cfg, fp_name=f"{fp_name}.act_fn"
+        )
+
+        # ----- local observers ----------------------------------------
+        self.act_in_obs = self._make_obs("act_in")
+        self.mul_obs = self._make_obs("mul")
+
+    def forward(self, x: torch.Tensor):
+        # 1) quantize input once
+        x_q = self._fq(x, self.act_in_obs)
+
+        # 2) parallel projections
+        g = self.gate_proj(x_q)
+        u = self.up_proj(x_q)
+
+        # 3) activation on gate
+        a = self.act_fn(g)
+
+        # 4) element-wise product
+        h = self._fq(a * u, self.mul_obs)
+
+        # 5) final projection
+        return self.down_proj(h)
+
+    def _all_observers(self):
+        # local first
+        yield self.act_in_obs
+        yield self.mul_obs
+        # recurse into children that are QuantModuleBase
+        for m in (self.gate_proj, self.up_proj, self.down_proj, self.act_fn):
+            yield from m._all_observers()

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

@@ -28,6 +28,8 @@ _CORE_MODULES = (
     "tico.experimental.quantization.ptq.wrappers.nn.quant_layernorm",
     "tico.experimental.quantization.ptq.wrappers.nn.quant_linear",
     "tico.experimental.quantization.ptq.wrappers.nn.quant_silu",
+    "tico.experimental.quantization.ptq.wrappers.llama.quant_attn",
+    "tico.experimental.quantization.ptq.wrappers.llama.quant_mlp",
     # add future core wrappers here
 )
 

{tico-0.1.0.dev250826.dist-info → tico-0.1.0.dev250828.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250826
+Version: 0.1.0.dev250828
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250826.dist-info → tico-0.1.0.dev250828.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=M4dQ4CTD_7xsO5DjUx76t4A5o1Q_2NqGlMe0fjkGDxQ,1883
+tico/__init__.py,sha256=Ur6T0ZsgBvl70ek5stJ0_uxhK9xNmHado-clOup1OEM,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
@@ -63,6 +63,8 @@ tico/experimental/quantization/ptq/qscheme.py,sha256=uwhv7bCxOOXB3I-IKlRyr_u4eXO
 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/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_mlp.py,sha256=N1qZQgt1S-xZrdv-PW7OfXEcv0gsO2q9faOF4aD-zKo,4147
 tico/experimental/quantization/ptq/observers/__init__.py,sha256=WF2MvL9M_jl-B1FqcY9zic34NOCRp17HkRYv-TMxMr4,613
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
 tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
@@ -76,7 +78,10 @@ tico/experimental/quantization/ptq/wrappers/__init__.py,sha256=47DEQpj8HBSa-_TIm
 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=562nKSlp9qF-w4-aQeJbx2V_wMGE2FRrjIKUfRwC4Mg,4571
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=TwH-MD-qkTkG6M-f1VqFLmSNcXLNYsh21yjyzCcojJc,4706
+tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=b360gkQ0RxExmiV-ZaaxwJdMPJ53g6uCRlR2-_dOby0,240
+tico/experimental/quantization/ptq/wrappers/llama/quant_attn.py,sha256=WIUI6EFMTvvruvqu8pBxWy6qJeDyjkaYbJk1R3pAmwE,8578
+tico/experimental/quantization/ptq/wrappers/llama/quant_mlp.py,sha256=uZMnrX66oZwxhKhcNbLXXeri-WxxRBiZnr15aBXJMm0,3562
 tico/experimental/quantization/ptq/wrappers/nn/__init__.py,sha256=I9uTt5HfcRoMEDYHpAeATMv2TbCQiX0ZbfUFMzSJ4Qw,336
 tico/experimental/quantization/ptq/wrappers/nn/quant_layernorm.py,sha256=G5Sgt-tXnzh0Rxyk-2honmZIfEQOZlRfOsoDBdSGmA4,6887
 tico/experimental/quantization/ptq/wrappers/nn/quant_linear.py,sha256=xW-VEPB7RJoslS3xLVCdhIuMjppknvpkZleRGK4JFVQ,2240
@@ -235,9 +240,9 @@ 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.dev250826.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250826.dist-info/METADATA,sha256=QhtUiHj_YT4ZxsClOx4OaP24kuaLUsLT83x3yl1gRDY,8450
-tico-0.1.0.dev250826.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250826.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250826.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250826.dist-info/RECORD,,
+tico-0.1.0.dev250828.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250828.dist-info/METADATA,sha256=nYk1Pl6H1eZbVSUDMUYLDjb8-AsRW9e9v1EAnIgLM4k,8450
+tico-0.1.0.dev250828.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250828.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250828.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250828.dist-info/RECORD,,