ByteDance

/

Ouro-1.4B-Thinking

like 31

Follow

ByteDance 4.28k

Problem

Batched generation with HuggingFace Transformers produces corrupted output for
all sequences except the longest (unpadded) one in the batch.

Root cause

UniversalTransformerCache inherits Cache.get_mask_sizes, which falls back to
return cache_position.shape[0], 0 when layer_idx >= len(self.layers).

Because UniversalTransformerCache manages its own flat key_cache /
value_cache lists and keeps self.layers empty ([]), this fallback always
fires. During the prefill step this happens to be correct (cache_position
spans the full input length), but during autoregressive decoding
cache_position has length 1, so the 4D attention mask is built for
kv_length=1 instead of cached_length + 1.

The undersized mask gets broadcasted across the full KV cache, losing all
per-position padding information. This corrupts every padded sequence in the
batch.

Fix

Override get_mask_sizes to return the correct (seq_length + query_length, 0),
matching the semantics of DynamicCacheLayer.get_mask_sizes.

Reproduction

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "ByteDance/Ouro-1.4B-Thinking"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto",
    attn_implementation="eager",
)

tokenizer.padding_side = "left"
tokenizer.pad_token_id = 0  # <|endoftext|>

# Two prompts of different lengths
prompts = ["What is 2+2?", "Explain why the sky is blue in one sentence."]
batch = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)

# Without fix: shorter prompt output is corrupted
outputs = model.generate(**batch, max_new_tokens=64, do_sample=False, eos_token_id=2, pad_token_id=0)
for i, p in enumerate(prompts):
    tokens = outputs[i][batch["input_ids"].shape[1]:]
    print(f"[{i}] {p!r} -> {tokenizer.decode(tokens, skip_special_tokens=False)[:100]}")