Instructions to use fxmeng/DSV2-Lite-IndexCache-Sparse-1250 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use fxmeng/DSV2-Lite-IndexCache-Sparse-1250 with Transformers:

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

pipe = pipeline("text-generation", model="fxmeng/DSV2-Lite-IndexCache-Sparse-1250", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("fxmeng/DSV2-Lite-IndexCache-Sparse-1250", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use fxmeng/DSV2-Lite-IndexCache-Sparse-1250 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "fxmeng/DSV2-Lite-IndexCache-Sparse-1250"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "fxmeng/DSV2-Lite-IndexCache-Sparse-1250",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/fxmeng/DSV2-Lite-IndexCache-Sparse-1250

SGLang

How to use fxmeng/DSV2-Lite-IndexCache-Sparse-1250 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 "fxmeng/DSV2-Lite-IndexCache-Sparse-1250" \
    --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": "fxmeng/DSV2-Lite-IndexCache-Sparse-1250",
		"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 "fxmeng/DSV2-Lite-IndexCache-Sparse-1250" \
        --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": "fxmeng/DSV2-Lite-IndexCache-Sparse-1250",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use fxmeng/DSV2-Lite-IndexCache-Sparse-1250 with Docker Model Runner:
```
docker model run hf.co/fxmeng/DSV2-Lite-IndexCache-Sparse-1250
```

DSV2-Lite-IndexCache-Sparse-1250 / validation.json

Model export

Add validated sparse step 1250 model aligned to archived DSA recipe

1e3cf35 4 days ago

Raw History Blame Contribute Delete

4.19 kB

	{
	"method": "indexcache",
	"num_tensors": 5336,
	"num_indexer_tensors": 45,
	"assignments": [
	{
	"layer": 0,
	"role": "group_leader",
	"leader": 0,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 1,
	"role": "shared",
	"leader": 0,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 2,
	"role": "shared",
	"leader": 0,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 3,
	"role": "group_leader",
	"leader": 3,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 4,
	"role": "shared",
	"leader": 3,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 5,
	"role": "shared",
	"leader": 3,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 6,
	"role": "group_leader",
	"leader": 6,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 7,
	"role": "shared",
	"leader": 6,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 8,
	"role": "shared",
	"leader": 6,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 9,
	"role": "group_leader",
	"leader": 9,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 10,
	"role": "shared",
	"leader": 9,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 11,
	"role": "shared",
	"leader": 9,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 12,
	"role": "group_leader",
	"leader": 12,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 13,
	"role": "shared",
	"leader": 12,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 14,
	"role": "shared",
	"leader": 12,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 15,
	"role": "group_leader",
	"leader": 15,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 16,
	"role": "shared",
	"leader": 15,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 17,
	"role": "shared",
	"leader": 15,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 18,
	"role": "group_leader",
	"leader": 18,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 19,
	"role": "shared",
	"leader": 18,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 20,
	"role": "shared",
	"leader": 18,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 21,
	"role": "group_leader",
	"leader": 21,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 22,
	"role": "shared",
	"leader": 21,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 23,
	"role": "shared",
	"leader": 21,
	"offset": 2,
	"group_size": 3
	},
	{
	"layer": 24,
	"role": "group_leader",
	"leader": 24,
	"offset": 0,
	"group_size": 3
	},
	{
	"layer": 25,
	"role": "shared",
	"leader": 24,
	"offset": 1,
	"group_size": 3
	},
	{
	"layer": 26,
	"role": "shared",
	"leader": 24,
	"offset": 2,
	"group_size": 3
	}
	],
	"all_tensor_shapes_match": true,
	"loading_info": {
	"missing_keys": [],
	"unexpected_keys": [],
	"mismatched_keys": [],
	"error_msgs": []
	},
	"device": "cuda:6",
	"decoder_cpu_offload": true,
	"cache_last_logit_max_abs_diff": 0.1875,
	"cache_last_logit_mean_abs_diff": 0.028124965727329254,
	"generation": "The capital of France is Paris, and the official language is French",
	"sparse_forward_length": 2057,
	"sparse_topk": 2048,
	"runtime_sha256": "712e109ef9d472dc4416def5601296c4f0bf6420f0dd8f7eacf634e6f2825460",
	"seconds": 213.86,
	"unit_tests_passed": 20,
	"unit_subtests_passed": 4,
	"all_exported_tensors_finite": true,
	"safetensors_index_consistent": true,
	"checkpoint_iteration": 1250,
	"validation_note": "GPU inference used decoder CPU offload because shared GPUs lacked free memory for all weights. Cache logit max-absolute BF16 tolerance: 0.5."
	}