Skip to content

fix(pi0-fast): don't apply embed scaling #3304

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
zucchini-nlp wants to merge 2 commits into
base: main
Choose a base branch
from
+7 −12
Open

fix(pi0-fast): don't apply embed scaling #3304

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3bfe179
weird i can't even pass the test on main
zucchini-nlp Apr 7, 2026
772c858
oops, typo
zucchini-nlp Apr 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 3 additions & 4 deletions src/lerobot/policies/pi0/modeling_pi0.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -443,13 +443,13 @@ def embed_image(self, image: torch.Tensor):
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
features = image_outputs.pooler_output
if features.dtype != out_dtype:
features = features.to(out_dtype)
return features

def embed_language_tokens(self, tokens: torch.Tensor):
return self.paligemma.model.language_model.embed_tokens(tokens)
return self.paligemma.model.language_model.get_input_embeddings()(tokens)

def forward(
self,
Expand Down Expand Up @@ -665,8 +665,7 @@ def image_embed_func(img):
# Process language tokens
def lang_embed_func(lang_tokens):
lang_emb = self.paligemma_with_expert.embed_language_tokens(lang_tokens)
lang_emb_dim = lang_emb.shape[-1]
return lang_emb * math.sqrt(lang_emb_dim)
return lang_emb

lang_emb = self._apply_checkpoint(lang_embed_func, lang_tokens)
embs.append(lang_emb)
Expand Down
12 changes: 4 additions & 8 deletions src/lerobot/policies/pi0_fast/modeling_pi0_fast.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,13 +258,13 @@ def embed_image(self, image: torch.Tensor):
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
features = image_outputs.pooler_output
if features.dtype != out_dtype:
features = features.to(out_dtype)
return features

def embed_language_tokens(self, tokens: torch.Tensor):
return self.paligemma.model.language_model.embed_tokens(tokens)
return self.paligemma.model.language_model.get_input_embeddings()(tokens)

def forward(
self,
Expand Down Expand Up @@ -414,8 +414,7 @@ def image_embed_func(img):
# Process language instruction tokens
def lang_embed_func(tokens):
lang_emb = self.paligemma_with_expert.embed_language_tokens(tokens)
lang_emb_dim = lang_emb.shape[-1]
return lang_emb * math.sqrt(lang_emb_dim)
return lang_emb
Comment on lines 414 to +417
Copy link

Copilot AI Apr 7, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

After removing the math.sqrt(...) scaling, this block no longer uses math; since the module is now unused in this file, the import math at the top should be removed to satisfy linters (ruff/flake8) and avoid dead imports.

Copilot uses AI. Check for mistakes.

lang_emb = self._apply_checkpoint(lang_embed_func, tokens)
embs.append(lang_emb)
Expand All @@ -429,8 +428,7 @@ def lang_embed_func(tokens):

def fast_action_embed_func(fast_action_tokens):
fast_emb = self.paligemma_with_expert.embed_language_tokens(fast_action_tokens)
fast_emb_dim = fast_emb.shape[-1]
return fast_emb * math.sqrt(fast_emb_dim)
return fast_emb

fast_action_emb = self._apply_checkpoint(fast_action_embed_func, fast_action_tokens)
embs.append(fast_action_emb)
Expand Down Expand Up @@ -663,7 +661,6 @@ def sample_actions_fast(
if t < max_decoding_steps - 1:
# embed the newly generated token
next_token_emb = self.paligemma_with_expert.embed_language_tokens(next_token)
next_token_emb = next_token_emb * math.sqrt(next_token_emb.shape[-1])
if prefix_embs.dtype == torch.bfloat16:
next_token_emb = next_token_emb.to(dtype=torch.bfloat16)

Expand Down Expand Up @@ -768,7 +765,6 @@ def sample_actions_fast_kv_cache(
# Embed the single previous token
# We use embed_language_tokens directly to avoid overhead of full prefix embedding
next_token_emb = self.paligemma_with_expert.embed_language_tokens(next_token)
next_token_emb = next_token_emb * math.sqrt(next_token_emb.shape[-1])
if prefix_embs.dtype == torch.bfloat16:
next_token_emb = next_token_emb.to(dtype=torch.bfloat16)

Expand Down
Loading