tico 0.2.0.dev260707__tar.gz → 0.2.0.dev260708__tar.gz

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tico
-Version: 0.2.0.dev260707
+Version: 0.2.0.dev260708
 Summary: Convert Exported Torch Module To Circle
 License: This file provides full text of licenses used in this project
         

tico-0.2.0.dev260708/tico/_version.py

@@ -0,0 +1 @@
+__version__ = "0.2.0.dev260708"

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/evaluation/vlm_eval_utils.py

@@ -273,7 +273,7 @@ DATASETS: dict[str, dict[str, Any]] = {
     "wikitext2": {
         "default_split": "test",
         "adapter": get_item_wikitext2,
-        "candidates": ["wikitext"],
+        "candidates": ["Salesforce/wikitext"],
         "config": "wikitext-2-raw-v1",
         "is_text_only": True,
     },

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/recipes/adapters/__init__.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from tico.quantization.recipes.adapters.base import ModelAdapter
+from tico.quantization.recipes.adapters.gemma4 import Gemma4Adapter
 from tico.quantization.recipes.adapters.llama import LlamaAdapter
 from tico.quantization.recipes.adapters.qwen3_vl import Qwen3VLAdapter
 
@@ -20,6 +21,7 @@ _ADAPTERS = {
     "llama": LlamaAdapter(),
     "qwen3_vl": Qwen3VLAdapter(),
     "qwen3-vl": Qwen3VLAdapter(),
+    "gemma4": Gemma4Adapter(),
 }
 
 

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/recipes/adapters/gemma4.py

@@ -23,6 +23,13 @@ from tico.quantization.config.gemma4_builders import build_gemma4_e2b_ptq_config
 from tico.quantization.recipes.adapters.base import ModelAdapter
 from tico.quantization.recipes.context import RecipeContext
 from tico.quantization.recipes.data.vlm import build_vlm_calibration_inputs
+from tico.quantization.recipes.evaluation.llava_bench_judge import (
+    evaluate_and_print_llava_bench_judge,
+)
+from tico.quantization.recipes.evaluation.vlm import (
+    evaluate_llava_bench,
+    print_coco_score_results,
+)
 from tico.quantization.recipes.utils import (
     move_to_device,
     quant_spec_from_config,
@@ -53,7 +60,12 @@ class Gemma4Adapter(ModelAdapter):
         cache_dir = model_cfg.get("cache_dir")
         device_map = runtime_cfg.get("device_map")
         if device_map is None:
-            device_map = "auto" if ctx.device.type != "cpu" else "cpu"
+            if ctx.device.type == "cpu":
+                device_map = "cpu"
+            else:
+                # Use specific device index instead of "auto" to avoid
+                # multi-GPU model split during PTQ calibration.
+                device_map = str(ctx.device)
 
         ctx.processor = AutoProcessor.from_pretrained(
             name,
@@ -174,11 +186,68 @@ class Gemma4Adapter(ModelAdapter):
         return convert(prepared, inplace=True)
 
     def evaluate(self, ctx: RecipeContext) -> None:
-        """Evaluate Gemma4 E2B.
+        """Evaluate Gemma4 E2B via LLaVA-Bench judge and other VLM tasks."""
+        eval_cfg = ctx.cfg.get("evaluation", {})
+        if not eval_cfg.get("enabled", False):
+            return
+
+        max_seq_len = eval_cfg.get("max_seq_len")
+        n_samples = int(eval_cfg.get("n_samples", 50))
+
+        llava_bench_cfg = eval_cfg.get("llava_bench", False)
+        if isinstance(llava_bench_cfg, Mapping):
+            if llava_bench_cfg.get("enabled", False):
+                mode = str(llava_bench_cfg.get("mode", "judge")).lower()
+                if mode in {"judge", "llm_judge"}:
+                    evaluate_and_print_llava_bench_judge(
+                        model=ctx.model,
+                        processor=ctx.processor,
+                        device=str(ctx.device),
+                        llava_cfg=llava_bench_cfg,
+                        model_cfg=ctx.cfg.get("model", {}),
+                        runtime_cfg=ctx.cfg.get("runtime", {}),
+                        default_n_samples=n_samples,
+                        default_max_seq_len=max_seq_len,
+                    )
+                elif mode in {"legacy", "coco", "caption"}:
+                    llava_results = evaluate_llava_bench(
+                        model=ctx.model,
+                        processor=ctx.processor,
+                        device=str(ctx.device),
+                        n_samples=int(llava_bench_cfg.get("n_samples", n_samples)),
+                        max_seq_len=llava_bench_cfg.get("max_seq_len", max_seq_len),
+                    )
+                    print_coco_score_results(
+                        "\n=== LLaVA Bench Legacy COCO-style Evaluation ===",
+                        llava_results,
+                    )
+                else:
+                    raise ValueError(
+                        "evaluation.llava_bench.mode must be one of "
+                        "{'judge', 'llm_judge', 'legacy', 'coco', 'caption'}, "
+                        f"got {mode!r}."
+                    )
+        elif llava_bench_cfg:
+            print(
+                "[WARNING] evaluation.llava_bench=true uses the legacy "
+                "COCO-style CIDEr/BLEU path. Prefer the nested judge config: "
+                "evaluation.llava_bench.enabled=true, mode=judge."
+            )
+            llava_results = evaluate_llava_bench(
+                model=ctx.model,
+                processor=ctx.processor,
+                device=str(ctx.device),
+                n_samples=n_samples,
+                max_seq_len=max_seq_len,
+            )
+            print_coco_score_results("\n=== Llava Bench Evaluation ===", llava_results)
+
+    def export(self, ctx: RecipeContext) -> None:
+        """Export Gemma4 E2B artifacts.
 
-        TODO: Reuse the Qwen3-VL LLaVA-Bench judge evaluator once the adapter is
-        wired into the shared evaluation entry point.
+        TODO: Implement Circle export via the static runtime submodule adapters.
         """
-        if not ctx.cfg.get("evaluation", {}).get("enabled", False):
+        export_cfg = ctx.cfg.get("export", {})
+        if not export_cfg.get("enabled", False):
             return
-        raise NotImplementedError("Gemma4 E2B evaluation adapter is not wired yet.")
+        raise NotImplementedError("Gemma4 E2B export adapter is not wired yet.")

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/recipes/adapters/llama.py

@@ -212,7 +212,7 @@ class LlamaAdapter(ModelAdapter):
             seq_len=seq_len,
             seed=int(runtime.get("seed", 42)),
             device=ctx.device,
-            dataset_name=calib.get("dataset", "wikitext"),
+            dataset_name=calib.get("dataset", "Salesforce/wikitext"),
             dataset_config=calib.get("dataset_config", "wikitext-2-raw-v1"),
             split=calib.get("split", "train"),
         )
@@ -326,7 +326,7 @@ class LlamaAdapter(ModelAdapter):
                 device=str(ctx.device),
                 cache_dir=ctx.cfg.get("model", {}).get("cache_dir"),
                 max_seq_len=max_seq_len,
-                dataset_name=ppl_cfg.get("dataset", "wikitext"),
+                dataset_name=ppl_cfg.get("dataset", "Salesforce/wikitext"),
                 dataset_config=ppl_cfg.get("dataset_config", "wikitext-2-raw-v1"),
                 split=ppl_cfg.get("split", "test"),
             )

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/recipes/data/llm.py

@@ -27,7 +27,7 @@ def build_wikitext_calibration_inputs(
     seq_len: int,
     seed: int,
     device: torch.device | str,
-    dataset_name: str = "wikitext",
+    dataset_name: str = "Salesforce/wikitext",
     dataset_config: str = "wikitext-2-raw-v1",
     split: str = "train",
 ) -> list[torch.Tensor]:

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/recipes/evaluation/llm.py

@@ -27,7 +27,7 @@ def evaluate_perplexity(
     device: str,
     cache_dir: str | None,
     max_seq_len: int,
-    dataset_name: str = "wikitext",
+    dataset_name: str = "Salesforce/wikitext",
     dataset_config: str = "wikitext-2-raw-v1",
     split: str = "test",
 ) -> float:

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/examples/compare_ppl.py

@@ -49,7 +49,7 @@ DTYPE_MAP = {
 }
 
 # Hardcoded dataset settings
-DATASET_NAME = "wikitext"
+DATASET_NAME = "Salesforce/wikitext"
 DATASET_CONFIG = "wikitext-2-raw-v1"
 TRAIN_SPLIT = "train"
 TEST_SPLIT = "test"

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/examples/debug_quant_outputs.py

@@ -65,7 +65,7 @@ DTYPE_MAP = {
 }
 
 # Hardcoded dataset settings
-DATASET_NAME = "wikitext"
+DATASET_NAME = "Salesforce/wikitext"
 DATASET_CONFIG = "wikitext-2-raw-v1"
 TRAIN_SPLIT = "train"
 

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/examples/quantize_full_qmodel_with_gptq.py

@@ -103,7 +103,7 @@ def _weight_dtype_from_bits(bits: int) -> DType:
 
 
 # Hardcoded dataset settings
-DATASET_NAME = "wikitext"
+DATASET_NAME = "Salesforce/wikitext"
 DATASET_CONFIG = "wikitext-2-raw-v1"
 TRAIN_SPLIT = "train"
 TEST_SPLIT = "test"
@@ -1037,13 +1037,14 @@ def get_sensitivities_info_name(model, dataset, seed, n_samples):
     Build a filename for stored sensitivity calibration results.
     """
     model_name = model.config.name_or_path.replace("/", "_")
+    dataset_name = dataset.replace("/", "_")
 
     name = (
         "."
         + "/sensitivities_for_"
         + model_name
         + "_"
-        + dataset
+        + dataset_name
         + "_"
         + str(n_samples)
         + "_"
@@ -1053,11 +1054,12 @@ def get_sensitivities_info_name(model, dataset, seed, n_samples):
     return name
 
 
-def get_ptq_model_name(model, args):
+def get_ptq_model_name(model, dataset, args):
     """
     Build a filename for a saved PTQ checkpoint.
     """
     model_name = model.config.name_or_path.replace("/", "_")
+    dataset_name = dataset.replace("/", "_")
 
     name = (
         f"PTQ_{model_name}_"
@@ -1393,7 +1395,7 @@ def save_requested_artifacts(q_m, tokenizer, calib_inputs, args) -> None:
     output_dir.mkdir(parents=True, exist_ok=True)
 
     if should_save(args, "ptq_checkpoint"):
-        save_path = output_dir / get_ptq_model_name(q_m.wrapped, args)
+        save_path = output_dir / get_ptq_model_name(q_m.wrapped, DATASET_NAME, args)
         print(f"Saving PTQ checkpoint to {save_path.resolve()}")
         torch.save(q_m, save_path)
 

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/examples/quantize_with_gptq.py

@@ -60,7 +60,7 @@ DTYPE_MAP = {
 }
 
 # Hardcoded dataset settings
-DATASET_NAME = "wikitext"
+DATASET_NAME = "Salesforce/wikitext"
 DATASET_CONFIG = "wikitext-2-raw-v1"
 TRAIN_SPLIT = "train"
 TEST_SPLIT = "test"

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/wrappers/gemma4/quant_model.py

@@ -93,10 +93,12 @@ class QuantGemma4Model(QuantModuleBase):
             qcfg=qcfg.child("embed_vision") if qcfg else None,
             fp_name=join_name(fp_name, "embed_vision"),
         )
-        if getattr(fp_model, "audio_tower", None) is not None:
-            raise NotImplementedError(
-                "Gemma4 E2B skeleton does not implement audio static runtime."
-            )
+        # Leave audio tower in FP (not quantized) - audio is not supported yet.
+        self.audio_tower = (
+            fp_model.audio_tower
+            if getattr(fp_model, "audio_tower", None) is not None
+            else None
+        )
 
         self.visual_start_idx = int(
             self.qcfg.model_args.get("vision", {}).get("visual_start_idx", 0)

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/wrappers/gemma4/quant_text_attention.py

@@ -20,7 +20,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from tico.quantization.config.ptq import PTQConfig
-from tico.quantization.wrapq.utils.utils import join_name
+from tico.quantization.wrapq.utils.utils import get_model_arg, join_name
 from tico.quantization.wrapq.wrappers.ptq_wrapper import PTQWrapper
 from tico.quantization.wrapq.wrappers.quant_module_base import QuantModuleBase
 from tico.quantization.wrapq.wrappers.registry import try_register
@@ -75,7 +75,7 @@ class QuantGemma4TextAttention(QuantModuleBase):
         self.attention_dropout = float(
             getattr(fp_attn, "attention_dropout", 0.0) or 0.0
         )
-        self.max_seq = int(getattr(self.config, "max_position_embeddings"))
+        self.max_seq = self._resolve_max_seq()
 
         self.q_proj = PTQWrapper(
             fp_attn.q_proj,
@@ -166,6 +166,53 @@ class QuantGemma4TextAttention(QuantModuleBase):
         mask.triu_(1)
         self.register_buffer("causal_mask_template", mask, persistent=False)
 
+    _DEFAULT_STATIC_MAX_SEQ = 2048
+
+    def _resolve_max_seq(self) -> int:
+        """Resolve the causal mask template capacity from config or model_args.
+
+        Resolution precedence:
+        1. ``calibration.seq_len`` (from PTQ config, same as Llama/Qwen adapters)
+        2. ``model_args["text"]["max_seq"]``
+        3. ``model_args["max_seq"]``
+        4. ``min(max_position_embeddings, 2048)``
+
+        This avoids allocating a ``(1, 1, max_position_embeddings, max_position_embeddings)``
+        buffer when the model supports a very large context window (e.g. 128K)
+        but the actual calibration/export sequence length is much smaller.
+        """
+        # Try calibration.seq_len first (same as Llama/Qwen adapters)
+        from tico.quantization.recipes.config import get_by_path
+
+        try:
+            calib_seq_len = get_by_path(self.qcfg, "calibration.seq_len")
+            if calib_seq_len is not None:
+                return int(calib_seq_len)
+        except (KeyError, TypeError):
+            pass
+
+        # Fall back to model_args
+        configured = get_model_arg(self.qcfg, "text", "max_seq", default=None)
+        if configured is None:
+            configured = get_model_arg(self.qcfg, "max_seq", default=None)
+
+        model_capacity = int(getattr(self.config, "max_position_embeddings"))
+        max_seq = (
+            min(model_capacity, self._DEFAULT_STATIC_MAX_SEQ)
+            if configured is None
+            else int(configured)
+        )
+        if max_seq <= 0:
+            raise ValueError(
+                f"Gemma4 attention max_seq must be positive, got {max_seq}."
+            )
+        if max_seq > model_capacity:
+            raise ValueError(
+                "Gemma4 attention max_seq exceeds max_position_embeddings: "
+                f"max_seq={max_seq}, model_capacity={model_capacity}."
+            )
+        return max_seq
+
     @staticmethod
     def _expand_rope_table(table: torch.Tensor) -> torch.Tensor:
         """Return a RoPE table shaped as ``(B, S, 1, H)`` for Gemma4 text."""

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/wrappers/gemma4/quant_vision_encoder.py

@@ -91,7 +91,9 @@ class QuantGemma4VisionEncoder(QuantModuleBase):
         attention_scaling = self.rotary_emb.attention_scaling
 
         max_pos = self.config.position_embedding_size  # e.g. 10240
-        position_indices = torch.arange(max_pos, dtype=torch.float)  # [0..max_pos-1]
+        position_indices = torch.arange(
+            max_pos, dtype=torch.float, device=inv_freq.device
+        )  # [0..max_pos-1]
         # freq_table[pos, i] = inv_freq[i] * pos  — equivalent to the matmul in original forward
         freq_table = torch.outer(
             position_indices, inv_freq

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico/quantization/wrapq/wrappers/gemma4/utils.py

@@ -132,9 +132,18 @@ def fixed_slot_fuse(
         )
 
     visual_len = int(visual_embeds.shape[1])
+    # Use actual visual token count - config num_visual_tokens is a hint for
+    # static export, but calibration images may produce different token counts.
     expected_len = visual_len if num_visual_tokens is None else int(num_visual_tokens)
     if visual_len != expected_len:
-        raise ValueError(f"Expected {expected_len} visual tokens, got {visual_len}.")
+        # Warn instead of raising error - use actual token count for flexibility
+        import warnings
+
+        warnings.warn(
+            f"Visual token count mismatch: expected {expected_len}, got {visual_len}. "
+            "Using actual count. If this is unexpected, check image resolution settings."
+        )
+        expected_len = visual_len
 
     end = int(visual_start_idx) + expected_len
     if visual_start_idx < 0 or end > seq_len:

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tico
-Version: 0.2.0.dev260707
+Version: 0.2.0.dev260708
 Summary: Convert Exported Torch Module To Circle
 License: This file provides full text of licenses used in this project
         

{tico-0.2.0.dev260707 → tico-0.2.0.dev260708}/tico.egg-info/SOURCES.txt

@@ -134,7 +134,7 @@ tico/quantization/evaluation/script/llm_tasks_eval.py
 tico/quantization/evaluation/script/mini_vqa_eval.py
 tico/quantization/examples/__init__.py
 tico/quantization/examples/evaluate.py
-tico/quantization/examples/inspect.py
+tico/quantization/examples/inspector.py
 tico/quantization/examples/quantize.py
 tico/quantization/examples/configs/__init__.py
 tico/quantization/passes/__init__.py

tico-0.2.0.dev260707/tico/_version.py

@@ -1 +0,0 @@
-__version__ = "0.2.0.dev260707"