Instructions to use InternScience/Sage-A1.5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use InternScience/Sage-A1.5 with Transformers:

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

pipe = pipeline("text-generation", model="InternScience/Sage-A1.5")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("InternScience/Sage-A1.5")
model = AutoModelForMultimodalLM.from_pretrained("InternScience/Sage-A1.5", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use InternScience/Sage-A1.5 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "InternScience/Sage-A1.5"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "InternScience/Sage-A1.5",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/InternScience/Sage-A1.5

SGLang

How to use InternScience/Sage-A1.5 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 "InternScience/Sage-A1.5" \
    --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": "InternScience/Sage-A1.5",
		"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 "InternScience/Sage-A1.5" \
        --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": "InternScience/Sage-A1.5",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use InternScience/Sage-A1.5 with Docker Model Runner:
```
docker model run hf.co/InternScience/Sage-A1.5
```

Sage-A1.5 compatible and optimized with Strata Fork - Tested on RTX 5080

by ykarout - opened 4 days ago

Discussion

ykarout

4 days ago

•

edited 3 days ago

Full model, Strata build and optimized engine on the repo below, now updated to round 2 (branch q36-r2):

ykarout/Strata-Sage-A1.5-GGUF

Round 2 adds tensor-core decode attention over the int8 cache, a fused prompt attention that reads the int8 cache directly, a chunked DeltaNet prompt pass, faster expert streaming and CPU-pool warm-up, and more experts kept in VRAM (78 -> 97 per layer at 262K). All of it is on by default, and each part can be switched off.

Measured

RTX 5080 16 GB, Core Ultra 9 285K, 64 GB DDR5-6400. Through the server: fresh code prompts, 384 tokens generated, temperature 0.7, thinking off, MTP 2, average of two runs, tok/s. llama.cpp b11834 tuned for this card: --cpu-moe --moe-cache-mib 10300, -b 4096 -ub 2048, MTP draft 2, 8 threads on the P-cores.

decode	4K	16K	40K	90K	160K	240K
Strata q36-r2 (262K context)	231	210	213	187	179	162
Strata q36-optimized (round 1)	173	154	151	124	105	86
llama.cpp (131K context)	166	171	146

prompt read	4K	16K	40K	90K	160K	240K
Strata q36-r2	6,425	6,903	6,115	4,561	3,351	2,477
Strata q36-optimized	4,685	5,111	4,350	2,706	1,511	806
llama.cpp	3,185	3,830	3,730

Decode +34% at 4K up to +88% at 240K over round 1, and 1.2-1.5x llama.cpp where both run. Prompt reads are 1.6-2.0x llama.cpp, and 3.1x round 1 at 240K. An agent's next turn on a cached 198K conversation (+2,000 tokens) takes 0.96 s (round 1: 5.1 s).

Quality. Round 2 changes numerics, so it was checked against the fp32 cache as the reference: every round-2 path is at least as close as round 1 (mean KL 6.6-8.5e-3 against 1.55e-2, full-vocabulary logits, 195 teacher-forced steps at 14K / 57K / 76K; top-1 184-189 of 195 against 185). At 200K the change is no bigger than changing the prompt chunk size. Needle retrieval at ~100K and ~200K: 40/40.

Sampling. The config on the model page sets your card's recommendation (temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence penalty 1.5) as the server default, because many agent clients (opencode, for one) send no temperature, which otherwise means greedy decoding.

Base Qwen3.6-35B-A3B on the same engine: ykarout/Strata-Qwen3.6-35B-A3B has the config, an expert profile and a how-to for Unsloth's MTP GGUFs (UD-Q4_K_XL recommended; the plain non-MTP repo does not load). Same card and sweep as above:

Qwen3.6-35B-A3B UD-Q4_K_XL	4K	16K	40K	90K	160K	240K
decode	228	211	199	186	167	159
prompt read	6,506	7,005	6,245	4,648	3,404	2,505

The base model accepts more MTP drafts (73-86% here, 88-95% on agent work), so in a real opencode coding session it decodes 273 tok/s under 16K context, 254 at 16-40K, 234 at 40-90K and 206 at 90-160K (peak 288).

The engine changes are proposed upstream to Strata as #3699 and #3903-#3911.

ykarout

3 days ago

repo is up to date now with the latest optimizations

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment