Instructions to use TaylorAI/Flash-Llama-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TaylorAI/Flash-Llama-7B with Transformers:

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

pipe = pipeline("text-generation", model="TaylorAI/Flash-Llama-7B", trust_remote_code=True)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("TaylorAI/Flash-Llama-7B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("TaylorAI/Flash-Llama-7B", trust_remote_code=True)

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TaylorAI/Flash-Llama-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TaylorAI/Flash-Llama-7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TaylorAI/Flash-Llama-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/TaylorAI/Flash-Llama-7B

SGLang

How to use TaylorAI/Flash-Llama-7B 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 "TaylorAI/Flash-Llama-7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TaylorAI/Flash-Llama-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "TaylorAI/Flash-Llama-7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TaylorAI/Flash-Llama-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use TaylorAI/Flash-Llama-7B with Docker Model Runner:
```
docker model run hf.co/TaylorAI/Flash-Llama-7B
```

andersonbcdefg commited on Aug 29, 2023

Commit

cc046e1

1 Parent(s): c5cda5b

Upload modeling_flash_llama.py

Browse files

Files changed (1) hide show

modeling_flash_llama.py +6 -1

modeling_flash_llama.py CHANGED Viewed

@@ -363,12 +363,17 @@ class LlamaAttention(nn.Module):
         past_key_value = (past_kv, past_len+q.size(1)) if use_cache else None
         # no padding tokens, more efficient
         attn_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
         attn_outputs = flash_attn_kvpacked_func(
             q.type(attn_dtype), kv.type(attn_dtype), dropout_p=0.0, softmax_scale=1.0/self.norm_factor, causal=(not has_layer_past), return_attn_probs=output_attentions)
         attn_output = attn_outputs[0] if output_attentions else attn_outputs
-        attn_output = attn_output.reshape(bsz, q_len, h_size)
         attn_weights = attn_outputs[2] if output_attentions else None
         if self.config.pretraining_tp > 1:

         past_key_value = (past_kv, past_len+q.size(1)) if use_cache else None
         # no padding tokens, more efficient
+        # the basic problem here is that for qlora, stuff is stored in float32, but attention needs float16 or bfloat16.
+        # if you cast just based on torch.cuda.is_bf_16_supported(), it works for training, but it breaks in evals
+        # that load the model in fp16 on a gpu where bf16 is supported. so, we cast based on the GPU support, but then
+        # cast back to whatever q originally was. hopefully that works!
+        orig_dtype = q.dtype
         attn_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
         attn_outputs = flash_attn_kvpacked_func(
             q.type(attn_dtype), kv.type(attn_dtype), dropout_p=0.0, softmax_scale=1.0/self.norm_factor, causal=(not has_layer_past), return_attn_probs=output_attentions)
         attn_output = attn_outputs[0] if output_attentions else attn_outputs
+        attn_output = attn_output.reshape(bsz, q_len, h_size).type(orig_dtype)
         attn_weights = attn_outputs[2] if output_attentions else None
         if self.config.pretraining_tp > 1: