refactor(smolvla): reuse shared VLA components (#4064)

* refactor(smolvla): reuse shared VLA components

* chore(policies): address review smolvla shared utilities

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -61,9 +61,15 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
 from lerobot.utils.constants import ACTION, OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS, OBS_STATE
-from lerobot.utils.device_utils import get_safe_dtype
 from lerobot.utils.import_utils import require_package
 
+from ..common.flow_matching import euler_integrate, sample_noise, sample_time_beta
+from ..common.vla_utils import (
+    create_sinusoidal_pos_embedding,
+    make_att_2d_masks,
+    pad_vector,
+    resize_with_pad,
+)
 from ..pretrained import PreTrainedPolicy
 from ..rtc.modeling_rtc import RTCProcessor
 from ..utils import (
@@ -79,96 +85,6 @@ class ActionSelectKwargs(TypedDict, total=False):
     execution_horizon: int | None
 
 
-def create_sinusoidal_pos_embedding(
-    time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
-) -> Tensor:
-    """Computes sine-cosine positional embedding vectors for scalar positions."""
-    if dimension % 2 != 0:
-        raise ValueError(f"dimension ({dimension}) must be divisible by 2")
-
-    if time.ndim != 1:
-        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
-
-    dtype = get_safe_dtype(torch.float64, device.type)
-    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
-    period = min_period * (max_period / min_period) ** fraction
-
-    # Compute the outer product
-    scaling_factor = 1.0 / period * 2 * math.pi
-    sin_input = scaling_factor[None, :] * time[:, None]
-    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-    return pos_emb
-
-
-def make_att_2d_masks(pad_masks, att_masks):
-    """Copied from big_vision.
-
-    Tokens can attend to valid inputs tokens which have a cumulative mask_ar
-    smaller or equal to theirs. This way `mask_ar` int[B, N] can be used to
-    setup several types of attention, for example:
-
-      [[1 1 1 1 1 1]]: pure causal attention.
-
-      [[0 0 0 1 1 1]]: prefix-lm attention. The first 3 tokens can attend between
-          themselves and the last 3 tokens have a causal attention. The first
-          entry could also be a 1 without changing behaviour.
-
-      [[1 0 1 0 1 0 0 1 0 0]]: causal attention between 4 blocks. Tokens of a
-          block can attend all previous blocks and all tokens on the same block.
-
-    Args:
-      input_mask: bool[B, N] true if its part of the input, false if padding.
-      mask_ar: int32[B, N] mask that's 1 where previous tokens cannot depend on
-        it and 0 where it shares the same attention mask as the previous token.
-    """
-    if att_masks.ndim != 2:
-        raise ValueError(att_masks.ndim)
-    if pad_masks.ndim != 2:
-        raise ValueError(pad_masks.ndim)
-
-    cumsum = torch.cumsum(att_masks, dim=1)
-    att_2d_masks = cumsum[:, None, :] <= cumsum[:, :, None]
-    pad_2d_masks = pad_masks[:, None, :] * pad_masks[:, :, None]
-    att_2d_masks = att_2d_masks & pad_2d_masks
-    return att_2d_masks
-
-
-def resize_with_pad(img, width, height, pad_value=-1):
-    # assume no-op when width height fits already
-    if img.ndim != 4:
-        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
-
-    cur_height, cur_width = img.shape[2:]
-
-    ratio = max(cur_width / width, cur_height / height)
-    resized_height = int(cur_height / ratio)
-    resized_width = int(cur_width / ratio)
-    resized_img = F.interpolate(
-        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
-    )
-
-    pad_height = max(0, int(height - resized_height))
-    pad_width = max(0, int(width - resized_width))
-
-    # pad on left and top of image
-    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
-    return padded_img
-
-
-def pad_vector(vector, new_dim):
-    """Can be (batch_size x sequence_length x features_dimension)
-    or (batch_size x features_dimension)
-    """
-    if vector.shape[-1] == new_dim:
-        return vector
-    shape = list(vector.shape)
-    current_dim = shape[-1]
-    shape[-1] = new_dim
-    new_vector = torch.zeros(*shape, dtype=vector.dtype, device=vector.device)
-    new_vector[..., :current_dim] = vector
-    return new_vector
-
-
 def normalize(x, min_val, max_val):
     return (x - min_val) / (max_val - min_val)
 
@@ -429,7 +345,13 @@ class SmolVLAPolicy(PreTrainedPolicy):
         for key in present_img_keys:
             img = batch[key][:, -1, :, :, :] if batch[key].ndim == 5 else batch[key]
             if self.config.resize_imgs_with_padding is not None:
-                img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
+                # SmolVLA stores the target as (width, height); the shared helper expects (height, width).
+                img = resize_with_pad(
+                    img,
+                    self.config.resize_imgs_with_padding[1],
+                    self.config.resize_imgs_with_padding[0],
+                    pad_value=0,
+                )
 
             # Normalize from range [0,1] to [-1,1] as expacted by siglip
             img = img * 2.0 - 1.0
@@ -619,20 +541,10 @@ class VLAFlowMatching(nn.Module):
             params.requires_grad = self.config.train_state_proj
 
     def sample_noise(self, shape, device):
-        noise = torch.normal(
+        return sample_noise(shape, device)
-            mean=0.0,
-            std=1.0,
-            size=shape,
-            dtype=torch.float32,
-            device=device,
-        )
-        return noise
 
     def sample_time(self, bsize, device):
-        beta_dist = torch.distributions.Beta(concentration1=1.5, concentration0=1.0)
+        return sample_time_beta(bsize, device, alpha=1.5, beta=1.0, scale=0.999, offset=0.001)
-        time_beta = beta_dist.sample((bsize,)).to(device=device, dtype=torch.float32)
-        time = time_beta * 0.999 + 0.001
-        return time
 
     def embed_prefix(
         self, images, img_masks, lang_tokens, lang_masks, state: torch.Tensor = None
@@ -800,7 +712,6 @@ class VLAFlowMatching(nn.Module):
             past_key_values=None,
             inputs_embeds=[prefix_embs, suffix_embs],
             use_cache=False,
-            fill_kv_cache=False,
         )
         suffix_out = suffix_out[:, -self.config.chunk_size :]
         # Original openpi code, upcast attention output
@@ -839,46 +750,24 @@ class VLAFlowMatching(nn.Module):
             past_key_values=None,
             inputs_embeds=[prefix_embs, None],
             use_cache=self.config.use_cache,
-            fill_kv_cache=True,
         )
         num_steps = self.config.num_steps
-        dt = -1.0 / num_steps
 
-        x_t = noise
+        return euler_integrate(
-        for step in range(num_steps):
+            lambda input_x_t, current_timestep: self.denoise_step(
-            time = 1.0 + step * dt
+                x_t=input_x_t,
-            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+                prefix_pad_masks=prefix_pad_masks,
-
+                past_key_values=past_key_values,
-            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                timestep=current_timestep,
-                return self.denoise_step(
+            ),
-                    x_t=input_x_t,
+            noise,
-                    prefix_pad_masks=prefix_pad_masks,
+            num_steps,
-                    past_key_values=past_key_values,
+            rtc_processor=self.rtc_processor,
-                    timestep=current_timestep,
+            rtc_enabled=self._rtc_enabled(),
-                )
+            inference_delay=kwargs.get("inference_delay"),
-
+            prev_chunk_left_over=kwargs.get("prev_chunk_left_over"),
-            if self._rtc_enabled():
+            execution_horizon=kwargs.get("execution_horizon"),
-                inference_delay = kwargs.get("inference_delay")
+        )
-                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-                execution_horizon = kwargs.get("execution_horizon")
-
-                v_t = self.rtc_processor.denoise_step(
-                    x_t=x_t,
-                    prev_chunk_left_over=prev_chunk_left_over,
-                    inference_delay=inference_delay,
-                    time=time,
-                    original_denoise_step_partial=denoise_step_partial_call,
-                    execution_horizon=execution_horizon,
-                )
-            else:
-                v_t = denoise_step_partial_call(x_t)
-
-            x_t = x_t + dt * v_t
-
-            if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
-                self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
-
-        return x_t
 
     def denoise_step(
         self,
@@ -907,8 +796,10 @@ class VLAFlowMatching(nn.Module):
             past_key_values=past_key_values,
             inputs_embeds=[None, suffix_embs],
             use_cache=self.config.use_cache,
-            fill_kv_cache=False,
         )
+        if past_key_values is not None:
+            # Self-attention layers append suffix K/V in place; restore the prefix for the next step.
+            past_key_values.crop(prefix_len)
         suffix_out = outputs_embeds[1]
         suffix_out = suffix_out[:, -self.config.chunk_size :]
         suffix_out = suffix_out.to(dtype=torch.float32)

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -26,6 +26,7 @@ if TYPE_CHECKING or _transformers_available:
         AutoModel,
         AutoModelForImageTextToText,
         AutoProcessor,
+        DynamicCache,
         SmolVLMForConditionalGeneration,
     )
 else:
@@ -33,6 +34,7 @@ else:
     AutoModel = None
     AutoModelForImageTextToText = None
     AutoProcessor = None
+    DynamicCache = None
     SmolVLMForConditionalGeneration = None
 
 
@@ -216,9 +218,8 @@ class SmolVLMWithExpertModel(nn.Module):
         batch_size,
         head_dim,
         use_cache: bool = True,
-        fill_kv_cache: bool = True,
+        past_key_values: "DynamicCache | None" = None,
-        past_key_values=None,
+    ) -> "tuple[list[torch.Tensor], DynamicCache | None]":
-    ) -> list[torch.Tensor]:
         query_states = []
         key_states = []
         value_states = []
@@ -259,22 +260,16 @@ class SmolVLMWithExpertModel(nn.Module):
         query_states = apply_rope(query_states, position_ids_)
         key_states = apply_rope(key_states, position_ids_)
 
-        if use_cache and past_key_values is None:
-            past_key_values = {}
-
         if use_cache:
-            if fill_kv_cache:
+            # `DynamicCache` stores tensors as [batch, heads, seq, head_dim]; this module works with
-                past_key_values[layer_idx] = {
+            # [batch, seq, heads, head_dim]. During prefix prefill this stores the (post-RoPE) K/V and
-                    "key_states": key_states,
+            # returns them unchanged; during denoising it appends the suffix K/V and returns
-                    "value_states": value_states,
+            # [prefix; suffix], exactly like the previous hand-rolled dict cache.
-                }
+            key_states, value_states = past_key_values.update(
-            else:
+                key_states.transpose(1, 2), value_states.transpose(1, 2), layer_idx
-                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
+            )
-                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
+            key_states = key_states.transpose(1, 2)
-                # the max len, then we (for instance) double the cache size. This implementation already exists
+            value_states = value_states.transpose(1, 2)
-                # in `transformers`. (molbap)
-                key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
-                value_states = torch.cat([past_key_values[layer_idx]["value_states"], value_states], dim=1)
 
         attention_interface = self.get_attention_interface()
 
@@ -293,13 +288,12 @@ class SmolVLMWithExpertModel(nn.Module):
         batch_size,
         head_dim,
         use_cache: bool = True,
-        fill_kv_cache: bool = True,
+        past_key_values: "DynamicCache | None" = None,
-        past_key_values=None,
+    ) -> "tuple[list[torch.Tensor], DynamicCache | None]":
-    ) -> list[torch.Tensor]:
         attention_interface = self.get_attention_interface()
 
         att_outputs = []
-        assert len(inputs_embeds) == 2 or (use_cache and past_key_values is not None and not fill_kv_cache), (
+        assert len(inputs_embeds) == 2 or (use_cache and past_key_values is not None), (
             f"Both len(inputs_embeds) == {len(inputs_embeds)} and past_key_values is {past_key_values}"
         )
 
@@ -332,22 +326,13 @@ class SmolVLMWithExpertModel(nn.Module):
         else:
             expert_position_id = position_ids
 
-        if use_cache and past_key_values is None:
+        if use_cache and past_key_values is not None:
-            past_key_values = {}
+            # Cross-attention layers never fill the cache themselves: during the prefix prefill every
-
+            # layer goes through `forward_attn_layer`, which stores the (post-RoPE) VLM K/V for this
-        if use_cache:
+            # layer index. Here we only read them back (no concatenation: the expert cross-attends to
-            if fill_kv_cache:
+            # the fixed prefix). `DynamicCache` stores [batch, heads, seq, head_dim]; transpose back.
-                past_key_values[layer_idx] = {
+            key_states = past_key_values.layers[layer_idx].keys.transpose(1, 2)
-                    "key_states": key_states,
+            value_states = past_key_values.layers[layer_idx].values.transpose(1, 2)
-                    "value_states": value_states,
-                }
-            else:
-                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
-                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
-                # the max len, then we (for instance) double the cache size. This implementation already exists
-                # in `transformers`. (molbap)
-                key_states = past_key_values[layer_idx]["key_states"]
-                value_states = past_key_values[layer_idx]["value_states"]
 
         # Expert
         expert_layer = model_layers[1][layer_idx]
@@ -360,14 +345,15 @@ class SmolVLMWithExpertModel(nn.Module):
             expert_hidden_states = expert_hidden_states.to(dtype=expert_layer.self_attn.q_proj.weight.dtype)
             expert_query_state = expert_layer.self_attn.q_proj(expert_hidden_states).view(expert_hidden_shape)
 
-            _key_states = key_states.to(dtype=expert_layer.self_attn.k_proj.weight.dtype).view(
+            # reshape (not view): K/V read back from the cache are transposed, hence non-contiguous
+            _key_states = key_states.to(dtype=expert_layer.self_attn.k_proj.weight.dtype).reshape(
                 *key_states.shape[:2], -1
             )
             expert_key_states = expert_layer.self_attn.k_proj(_key_states).view(
                 *_key_states.shape[:-1], -1, expert_layer.self_attn.head_dim
             )  # k_proj should have same dim as kv
 
-            _value_states = value_states.to(dtype=expert_layer.self_attn.v_proj.weight.dtype).view(
+            _value_states = value_states.to(dtype=expert_layer.self_attn.v_proj.weight.dtype).reshape(
                 *value_states.shape[:2], -1
             )
             expert_value_states = expert_layer.self_attn.v_proj(_value_states).view(
@@ -416,10 +402,9 @@ class SmolVLMWithExpertModel(nn.Module):
         self,
         attention_mask: torch.Tensor | None = None,
         position_ids: torch.LongTensor | None = None,
-        past_key_values: list[torch.FloatTensor] | None = None,
+        past_key_values: "DynamicCache | None" = None,
         inputs_embeds: list[torch.FloatTensor] = None,
         use_cache: bool | None = None,
-        fill_kv_cache: bool | None = None,
     ):
         models = [self.get_vlm_model().text_model, self.lm_expert]
         model_layers = self.get_model_layers(models)
@@ -431,6 +416,13 @@ class SmolVLMWithExpertModel(nn.Module):
                 continue
             batch_size = hidden_states.shape[0]
 
+        # Prefix prefill: no cache was passed, so create one and fill it (every layer runs
+        # self-attention over the prefix). When a filled cache is passed (denoising), layers
+        # read from it instead.
+        fill_kv_cache = use_cache and past_key_values is None
+        if fill_kv_cache:
+            past_key_values = DynamicCache()
+
         # RMSNorm
         num_layers = self.num_vlm_layers
         head_dim = self.vlm.config.text_config.head_dim
@@ -449,7 +441,6 @@ class SmolVLMWithExpertModel(nn.Module):
                     batch_size,
                     head_dim,
                     use_cache=use_cache,
-                    fill_kv_cache=fill_kv_cache,
                     past_key_values=past_key_values,
                 )
             else:
@@ -462,7 +453,6 @@ class SmolVLMWithExpertModel(nn.Module):
                     batch_size,
                     head_dim,
                     use_cache=use_cache,
-                    fill_kv_cache=fill_kv_cache,
                     past_key_values=past_key_values,
                 )
             outputs_embeds = []