[showo2] Fix env setup and T2I generation failure when resolution > 432

[BlueDyee]

Problem

This PR fixes two critical issues that prevent inference at resolutions different from the pretrained model's training resolution (432x432):

1. Dependency Conflict in Environment Setup

Issue: Installing flash-attn overwrites PyTorch version

• torch==2.5.1 is installed first
• flash-attn installation downgrades to torch==2.9.0 (incompatible)

Root Cause: flash-attn has strict PyTorch version requirements that conflict with the specified version.

Solution: Update requirements.txt with compatible versions and installation order.

2. Position Embedding Dimension Mismatch

Issue: RuntimeError when using resolutions different from pretrained model

RuntimeError: The size of tensor a (1024) must match the size of tensor b (729) at non-singleton dimension 1

Root Cause:

• Pretrained model config (show-o2-1.5B-HQ/config.json) contains fixed resolution settings:

  "image_latent_height": 27,
  "image_latent_width": 27,

• __init__ loads model with these hardcoded values (27×27=729 tokens)

  self.register_buffer("image_position_ids",
                     torch.arange(image_latent_height * image_latent_width).expand((1, -1)),
                     persistent=False)
  

• When inferring at 1024×1024 resolution, actual input has 64×64=4096 tokens
• The image_position_ids buffer (initialized at Line 81) remains at 729 tokens
• Position embedding fails when trying to add embeddings of different sizes

Solution:
Add dynamic position ID recreation in the forward() method to match actual input shape:

# Check if actual input shape matches buffered position IDs
actual_seq_len = h_ * w_
if actual_seq_len != self.image_position_ids.shape[-1]:
    # Dynamically recreate position IDs to match input
    self.register_buffer(
        "image_position_ids",
        torch.arange(actual_seq_len, device=self.image_position_ids.device).expand((1, -1)),
        persistent=False
    )

This allows the model to handle any input resolution dynamically, utilizing the existing position embedding interpolation logic when needed.

Changes

Files Modified

1. requirements.txt (NEW)

   • Add comprehensive dependency list with compatible versions
   • Specify installation order to avoid conflicts
   • Include notes for flash-attn installation

2. models/modeling_showo2_qwen2_5.py

   • Add dynamic image_position_ids recreation in forward() method (Line ~310)
   • Ensures position embeddings match actual input dimensions
   • Preserves existing interpolation logic for non-standard resolutions

Testing

Tested configurations:

• ✅ 432×432 (original resolution)
• ✅ 1024×1024