deepseek-ai

/

DeepSeek-OCR

like 2.43k

Follow

DeepSeek 104k

Issue

The model fails to load with new Transformers versions due to removed classes:

ImportError: cannot import name 'LlamaFlashAttention2' from 'transformers.models.llama.modeling_llama'

Root Cause

In modeling_deepseekv2.py (lines 37-39), the code imports:

from transformers.models.llama.modeling_llama import (
    LlamaAttention,
    LlamaFlashAttention2
)

These classes were removed in Transformers 4.47+ as part of the attention refactoring.

Proposed Fix

Since DeepSeek-OCR uses MLA (Multi-head Latent Attention) by default (config.use_mla = True), the Llama attention classes are only used as fallbacks for MHA mode.

Option 1: Remove MHA support (simplest)

1. Remove the imports (lines 37-39)
2. Update ATTENTION_CLASSES dict (lines 1022-1029):

ATTENTION_CLASSES = {
    "eager": DeepseekV2Attention,
    "flash_attention_2": DeepseekV2FlashAttention2,
    "mla_eager": DeepseekV2Attention,
    "mla_flash_attention_2": DeepseekV2FlashAttention2,
    # Removed mha_eager and mha_flash_attention_2
}

Option 2: Use DeepSeek attention for MHA mode (backward compatible)

Keep the same keys but map to DeepSeek classes:

ATTENTION_CLASSES = {
    "eager": DeepseekV2Attention,
    "flash_attention_2": DeepseekV2FlashAttention2,
    "mla_eager": DeepseekV2Attention,
    "mla_flash_attention_2": DeepseekV2FlashAttention2,
    "mha_eager": DeepseekV2Attention,  # Changed
    "mha_flash_attention_2": DeepseekV2FlashAttention2,  # Changed
}

Option 3: Conditional import (most flexible)

try:
    from transformers.models.llama.modeling_llama import (
        LlamaAttention,
        LlamaFlashAttention2
    )
    HAS_LLAMA_ATTENTION = True
except ImportError:
    HAS_LLAMA_ATTENTION = False

ATTENTION_CLASSES = {
    "eager": DeepseekV2Attention,
    "flash_attention_2": DeepseekV2FlashAttention2,
    "mla_eager": DeepseekV2Attention,
    "mla_flash_attention_2": DeepseekV2FlashAttention2,
}

if HAS_LLAMA_ATTENTION:
    ATTENTION_CLASSES.update({
        "mha_eager": LlamaAttention,
        "mha_flash_attention_2": LlamaFlashAttention2
    })
else:
    ATTENTION_CLASSES.update({
        "mha_eager": DeepseekV2Attention,
        "mha_flash_attention_2": DeepseekV2FlashAttention2
    })

This works because DeepSeek-OCR uses MLA by default anyway.