chore(dependencies): bump transformers v5 (#2964)

* chore(dependencies): upgrade transformers + hggingface-hub + peft + scipy

* chore(dependencies): bump pi0 family to transformers v5

* chore(dependencies): bump wall x to transformers v5

* chore(dependencies): bump gr00t to transformers v5

* chore(style): fix pre-commit

* fix(policy): xvla forced_bos_token missing

* test(rl): skip ci tests for resnet10

* Fix: full pi models support for transformer v5 (#2967)

* fix(pi): remove loss truncation

* fix(pi): remove state padding before tokenization

* fix(pi): fix image padding value

* fix from_pretrain

* add transformer v5 changes

* remove reference

* more fixes

* make it work

* add support for rest of pi family

* add pifast work

* more changes

* more changes

* more cleanup

* fix torch params

* dtype fix

* torch compile

* embed mismatch fix

* revert groot

* more nit fixes

* remove unused classes

* more fixes

* revert

* nit

* torch dtype warning fix

* but back dynamic renaming

* add tie embedding

---------

Co-authored-by: Yufei Sun <skieyfly@gmail.com>

* chore: fix XVLA in transformers v5 (#3006)

* test(policies): enable wall x CI testing

* style(test): pre-commit check

* style(test): pre-commit

* fix wall x for transformer v5 (#3008)

* tv5 fix

* various wall x fixes

* Delete tests/policies/pi0_pi05/print_pi05_output_logits.py

Signed-off-by: Jade Choghari <chogharijade@gmail.com>

* sync modeling_florence2.py with chore/bump_transformers_v5

* more

* more fixes

* more

* remove comment

* more

---------

Signed-off-by: Jade Choghari <chogharijade@gmail.com>

* chore(dependencies): adjust dependencies versioning after transformers v5 (#3034)

* chore(dependecies): adjust dependecies versioning after transformers v5

* fix(policies): remove deprecated input_embeds

* fix(policies): dict _tied_weights_keys

* chore(depedencies): common qwen-vl-utils

* chore(dependencies): bump transformers to 5.2

* Fix policy testing for tv5 (#3032)

* fix ci logger

* other fix

* fix mypy

* change logits to torch2.10

* skip wallx|

* remove logging

---------

Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

* feat(ci): log into HF to unblock some CI tests (#3007)

* feat(ci): log into HF to unblock some CI tests

* chore(ci): change hf call + secret name

* fix(ci): temp fix for pi0 rtc test

* test(policies): require_cuda for unblocked tests

* test(policies): require_cuda wall_x

* fic(tests): require_cuda outter most for pi0

* fix(test): return instead of yield

---------

Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>

* style(test): fix pre-commit

* chore(deps): upgrade transformers (#3050)

* chore(test): use lerobot model

* fix(policies): change default action tokenizer for wall x

* sample on cpu

* Revert "Merge branch 'chore/bump_transformers_v5' of https://github.com/huggingface/lerobot into chore/bump_transformers_v5"

This reverts commit d9b76755f7, reversing
changes made to 89359cb0b6.

* Reapply "Merge branch 'chore/bump_transformers_v5' of https://github.com/huggingface/lerobot into chore/bump_transformers_v5"

This reverts commit c9914db78b.

---------

Signed-off-by: Jade Choghari <chogharijade@gmail.com>
Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: Jade Choghari <chogharijade@gmail.com>
Co-authored-by: Yufei Sun <skieyfly@gmail.com>
Co-authored-by: Pepijn <pepijn@huggingface.co>

src/lerobot/policies/pi05/modeling_pi05.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
@@ -32,14 +33,20 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
-    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
-
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
@@ -92,10 +99,11 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
-    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
+    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
+    alpha_t = torch.tensor(alpha, dtype=torch.float32)
+    beta_t = torch.tensor(beta, dtype=torch.float32)
     dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,))
+    return dist.sample((bsize,)).to(device)
 
 
 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -189,7 +197,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -200,7 +208,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -219,14 +227,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
     layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.language_model, gemma_expert.model]
+    models = [paligemma.model.language_model, gemma_expert.model]
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
         layer = models[i].layers[layer_idx]
-        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -252,10 +260,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -263,7 +271,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
@@ -275,15 +283,15 @@ def compute_layer_complete(
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
         out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
         # first residual
-        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
         # Convert to bfloat16 if the next layer (mlp) uses bfloat16
         if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
             out_emb = out_emb.to(dtype=torch.bfloat16)
         out_emb = layer.mlp(out_emb)
         # second residual
-        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -356,7 +364,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -364,7 +372,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -375,13 +383,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -396,10 +404,11 @@ class PaliGemmaWithExpertModel(
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Saves memory vs full float32; more memory than only 3 params.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -411,8 +420,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -422,15 +431,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32).
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -444,7 +461,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -468,7 +485,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.language_model, self.gemma_expert.model]
             num_layers = self.paligemma.config.text_config.num_hidden_layers
 
             # Check if gradient checkpointing is enabled for any of the models
@@ -508,7 +525,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -573,29 +590,19 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Also compile the main forward pass used during training
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI05Pytorch model")
 
     def gradient_checkpointing_disable(self):
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
@@ -737,7 +744,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(x_t, time)
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -808,7 +815,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
         prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -880,6 +887,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
+        past_key_values = copy.deepcopy(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,
@@ -969,14 +977,12 @@ class PI05Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -984,7 +990,7 @@ class PI05Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -997,7 +1003,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
 
             # Then add "model." prefix for all keys that don't already have it
@@ -1009,8 +1015,6 @@ class PI05Policy(PreTrainedPolicy):
                     new_key = f"model.{key}"
                     remapped_state_dict[new_key] = value
                     remap_count += 1
-                    if remap_count <= 10:  # Only print first 10 to avoid spam
-                        print(f"Remapped: {key} -> {new_key}")
                 else:
                     remapped_state_dict[key] = value
 
@@ -1044,7 +1048,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1098,6 +1102,14 @@ class PI05Policy(PreTrainedPolicy):
                 # Some checkpoints might have this, but current model expects different structure
                 logging.warning(f"Vision embedding key might need handling: {key}")
 
+            if (
+                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
+                or key == "paligemma_with_expert.paligemma.lm_head.weight"
+            ):
+                fixed_state_dict[
+                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
+                ] = value.clone()
+
             fixed_state_dict[new_key] = value
 
         return fixed_state_dict

src/lerobot/policies/pi05/processor_pi05.py

@@ -23,7 +23,6 @@ import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
-from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -68,9 +67,6 @@ class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
         # TODO: check if this necessary
         state = deepcopy(state)
 
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
         # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
         # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
         state_np = state.cpu().numpy()