Instructions to use ByteDance-Seed/Stable-DiffCoder-8B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ByteDance-Seed/Stable-DiffCoder-8B-Instruct with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ByteDance-Seed/Stable-DiffCoder-8B-Instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("ByteDance-Seed/Stable-DiffCoder-8B-Instruct", trust_remote_code=True)
model = AutoModel.from_pretrained("ByteDance-Seed/Stable-DiffCoder-8B-Instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use ByteDance-Seed/Stable-DiffCoder-8B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ByteDance-Seed/Stable-DiffCoder-8B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ByteDance-Seed/Stable-DiffCoder-8B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct

SGLang

How to use ByteDance-Seed/Stable-DiffCoder-8B-Instruct with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ByteDance-Seed/Stable-DiffCoder-8B-Instruct" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ByteDance-Seed/Stable-DiffCoder-8B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ByteDance-Seed/Stable-DiffCoder-8B-Instruct" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ByteDance-Seed/Stable-DiffCoder-8B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ByteDance-Seed/Stable-DiffCoder-8B-Instruct with Docker Model Runner:
```
docker model run hf.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct
```

Natt1e commited on Mar 15

Commit

e71e35b

verified ·

1 Parent(s): bc14582

fix bug for generaion(empty response and PyTorch exception)

Browse files

This PR is the same as the https://github.com/ByteDance-Seed/Stable-DiffCoder/issues/1

It fixes two bugs:
1. empty response

2. PyTorch exception

Files changed (1) hide show

modeling_stable_diffcoder.py +21 -9

modeling_stable_diffcoder.py CHANGED Viewed

@@ -137,8 +137,10 @@ class StableDiffcoderForCausalLM(LlamaForCausalLM):
         prompt_length = input_ids.shape[1]
         gen_block_list = [block_length for _ in range(gen_blocks)]
-        res_block = block_length - (prompt_length % block_length)
-        if res_block > 0:
             gen_block_list = [res_block] + gen_block_list
             gen_block_list[-1] = block_length - res_block
             gen_blocks += 1
@@ -156,16 +158,20 @@ class StableDiffcoderForCausalLM(LlamaForCausalLM):
         nfe = 0
         final_flag = False
         prefill_length = prompt_length // block_length * block_length
         if prefill_length > 0:
             cur_attn_mask = block_diffusion_attention_mask[
                 ..., :prefill_length, :prefill_length
             ]
             self(
                 x[:, :prefill_length],
                 past_key_values=past_key_values,
                 attention_mask=cur_attn_mask,
                 use_cache=True,
-            ).past_key_values
         for block_id, block_size in enumerate(gen_block_list):
             block_start = (
@@ -182,7 +188,7 @@ class StableDiffcoderForCausalLM(LlamaForCausalLM):
             replace_position[:, block_start:block_end] = True
             for token_count in num_transfer_tokens:
-                if token_count:
                     nfe += 1
                     mask_map = x[:, block_start:block_end] == mask_id
                     attention_mask = block_diffusion_attention_mask[
@@ -205,22 +211,28 @@ class StableDiffcoderForCausalLM(LlamaForCausalLM):
                         remasking,
                         mask_map,
                         x[:, block_start:block_end],
-                        token_count if threshold is None else None,
                         threshold,
-                        shift=False,
                     )
                     x[:, block_start:block_end][transfer_map] = x0[transfer_map]
                 if (x[:, block_start:block_end] == mask_id).sum() == 0:
                     if (
                         eos_id is not None
-                        and (x[:, block_start:block_end] == eos_id).sum() > 0
                     ):
                         final_flag = True
                         x = x[:, :block_end]
-                        eos_pos = (x == eos_id).nonzero(as_tuple=True)[1][0].item()
                         x[0, eos_pos:] = eos_id
                         break
                     nfe += 1
                     self(
                         x[:, block_start:block_end],
@@ -231,7 +243,7 @@ class StableDiffcoderForCausalLM(LlamaForCausalLM):
                         use_cache=True,
                         cache_position=replace_position.nonzero(as_tuple=True)[1],
                     )
-                    break
             if final_flag:
                 break

         prompt_length = input_ids.shape[1]
         gen_block_list = [block_length for _ in range(gen_blocks)]
+        # Fix 3: Only handle residual blocks if the prompt length is NOT cleanly divisible
+        remainder = prompt_length % block_length
+        if remainder != 0:
+            res_block = block_length - remainder
             gen_block_list = [res_block] + gen_block_list
             gen_block_list[-1] = block_length - res_block
             gen_blocks += 1
         nfe = 0
         final_flag = False
         prefill_length = prompt_length // block_length * block_length
         if prefill_length > 0:
             cur_attn_mask = block_diffusion_attention_mask[
                 ..., :prefill_length, :prefill_length
             ]
+            # Fix 1: Explicitly pass cache_position for newer transformers prefill
+            cache_pos = torch.arange(prefill_length, device=x.device)
             self(
                 x[:, :prefill_length],
                 past_key_values=past_key_values,
                 attention_mask=cur_attn_mask,
                 use_cache=True,
+                cache_position=cache_pos,
+            )
         for block_id, block_size in enumerate(gen_block_list):
             block_start = (
             replace_position[:, block_start:block_end] = True
             for token_count in num_transfer_tokens:
+                if token_count > 0:
                     nfe += 1
                     mask_map = x[:, block_start:block_end] == mask_id
                     attention_mask = block_diffusion_attention_mask[
                         remasking,
                         mask_map,
                         x[:, block_start:block_end],
+                        token_count.item() if threshold is None else None,
                         threshold,
+                        shift=shift,
                     )
                     x[:, block_start:block_end][transfer_map] = x0[transfer_map]
                 if (x[:, block_start:block_end] == mask_id).sum() == 0:
+                   # Fix 2: Calculate where the generated tokens ACTUALLY start in this block
+                    gen_start = max(block_start, prompt_length)
                     if (
                         eos_id is not None
+                        and gen_start < block_end
+                        and (x[:, gen_start:block_end] == eos_id).sum() > 0
                     ):
                         final_flag = True
                         x = x[:, :block_end]
+                        eos_pos = (x[:, gen_start:block_end] == eos_id).nonzero(as_tuple=True)[1][0].item() + gen_start
                         x[0, eos_pos:] = eos_id
                         break
                     nfe += 1
                     self(
                         x[:, block_start:block_end],
                         use_cache=True,
                         cache_position=replace_position.nonzero(as_tuple=True)[1],
                     )
+                    break
             if final_flag:
                 break