Instructions to use sarvamai/sarvam-105b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sarvamai/sarvam-105b with Transformers:

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

pipe = pipeline("text-generation", model="sarvamai/sarvam-105b", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("sarvamai/sarvam-105b", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use sarvamai/sarvam-105b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sarvamai/sarvam-105b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sarvamai/sarvam-105b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sarvamai/sarvam-105b

SGLang

How to use sarvamai/sarvam-105b 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 "sarvamai/sarvam-105b" \
    --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": "sarvamai/sarvam-105b",
		"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 "sarvamai/sarvam-105b" \
        --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": "sarvamai/sarvam-105b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sarvamai/sarvam-105b with Docker Model Runner:
```
docker model run hf.co/sarvamai/sarvam-105b
```

Fix MoE routing and training gradients

#13

by tejesh-sarvam - opened 20 days ago

base: refs/heads/main

←

from: refs/pr/13

Discussion Files changed

+31

-22

Fix MoE routing and training gradientsa7aa449e

tejesh-sarvam

20 days ago

•

edited 16 days ago

Summary

match vLLM routing by always using global correction-biased top-k for Sarvam-105B
reject n_group or topk_group configuration with a clear error because grouped routing is not supported by this patch
keep e_score_correction_bias as persistent checkpoint state without treating it as a trainable parameter
restore differentiable single-rank MoE dispatch during training and fail clearly for unsupported expert-parallel training

Validation

NVIDIA B200, fixed seed 20260911, vLLM 0.28.0
production-shaped router: 128 experts, top-8, 51 tokens; expert IDs matched exactly and maximum routing-weight absolute error was 5.96e-08
miniature two-layer Sarvam checkpoint loaded independently by Transformers and vLLM: identical greedy tokens [127, 200, 155, 155]; maximum absolute error over 1,024 vocabulary log-probabilities was 0.0019813
CPU training smoke test confirmed finite gradients for inputs, router weights, and expert parameters

Disable unsupported grouped MoE routing35b2d7f4

Remove local MoE regression test1a02864d

rahular changed pull request status to merged 10 days ago

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