Instructions to use saricles/MiniMax-M2.7-NVFP4-GB10-AC with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use saricles/MiniMax-M2.7-NVFP4-GB10-AC with Transformers:

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

pipe = pipeline("text-generation", model="saricles/MiniMax-M2.7-NVFP4-GB10-AC", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("saricles/MiniMax-M2.7-NVFP4-GB10-AC", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("saricles/MiniMax-M2.7-NVFP4-GB10-AC", trust_remote_code=True, device_map="auto")
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 Settings

vLLM

How to use saricles/MiniMax-M2.7-NVFP4-GB10-AC with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "saricles/MiniMax-M2.7-NVFP4-GB10-AC"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "saricles/MiniMax-M2.7-NVFP4-GB10-AC",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/saricles/MiniMax-M2.7-NVFP4-GB10-AC

SGLang

How to use saricles/MiniMax-M2.7-NVFP4-GB10-AC 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 "saricles/MiniMax-M2.7-NVFP4-GB10-AC" \
    --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": "saricles/MiniMax-M2.7-NVFP4-GB10-AC",
		"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 "saricles/MiniMax-M2.7-NVFP4-GB10-AC" \
        --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": "saricles/MiniMax-M2.7-NVFP4-GB10-AC",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use saricles/MiniMax-M2.7-NVFP4-GB10-AC with Docker Model Runner:
```
docker model run hf.co/saricles/MiniMax-M2.7-NVFP4-GB10-AC
```

MiniMax-M2.7-NVFP4-GB10-AC / run_vllm.sh

saricles

run_vllm.sh: default to Marlin NVFP4 MoE + ngram (Agentic profile); SPECULATIVE_CONFIG var; rationale comments

0ba0662 verified 4 months ago

Raw

History Blame Contribute Delete

4.1 kB

	#!/usr/bin/env bash
	# run_vllm.sh — reference vLLM launch for MiniMax-M2.7-NVFP4-GB10-AC on dual-Spark TP=2.
	#
	# Ships configured for the "Agentic" deployment profile (Marlin NVFP4 MoE + ngram speculative
	# decoding). Comment out the --speculative-config line below to switch to the
	# "Throughput-stable" profile for novel-text / batch workloads.
	#
	# See DEPLOYMENT.md in this repo for:
	# - Profile tradeoffs and when to pick which
	# - Measured numbers on 2× DGX Spark (GB10, SM 12.1)
	# - Observations, caveats, and links to the community threads / PRs that informed this recipe.
	#
	# Assumes:
	# - Ray head + worker already running (one per Spark)
	# - Model mounted/available at $MODEL_PATH on both hosts
	# - vLLM >= 0.19.x with the Marlin NVFP4 backend built in (eugr/spark-vllm-docker nightly is the reference image)

	set -euo pipefail

	MODEL_PATH="${MODEL_PATH:-/models/MiniMax-M2.7-NVFP4-GB10-AC}"
	SERVED_NAME="${SERVED_NAME:-minimax-m2.7-ac}"
	HOST="${HOST:-0.0.0.0}"
	PORT="${PORT:-30000}"
	GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.88}"
	MAX_MODEL_LEN="${MAX_MODEL_LEN:-196608}"
	MAX_NUM_SEQS="${MAX_NUM_SEQS:-12}"
	MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-32768}"
	TP_SIZE="${TP_SIZE:-2}"

	# --- Tuned environment variables ----------------------------------------------
	# Forum + vendor-recipe validated for MiniMax-M2.7 NVFP4 on GB10 (SM 12.1).
	# On SM 12.1, the Marlin NVFP4 MoE backend is currently the fastest path — the
	# FlashInfer CUTLASS NVFP4 MoE path has maturity issues on this specific compute
	# capability (see DEPLOYMENT.md § "Why Marlin MoE on GB10").

	export SAFETENSORS_FAST_GPU=1
	export OMP_NUM_THREADS=8
	export TORCHINDUCTOR_MAX_AUTOTUNE=0

	export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
	export VLLM_FLOAT32_MATMUL_PRECISION=high
	export VLLM_FLASHINFER_MOE_BACKEND=throughput
	export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1

	# Marlin NVFP4 MoE path
	export VLLM_NVFP4_GEMM_BACKEND=marlin
	export VLLM_USE_FLASHINFER_MOE_FP4=0
	export VLLM_TEST_FORCE_FP8_MARLIN=1
	export VLLM_MARLIN_USE_ATOMIC_ADD=1

	# --- Compilation config -------------------------------------------------------
	# cudagraph_mode=none is INTENTIONAL on dual-Spark Ray TP.
	# PIECEWISE captures cleanly in current vLLM builds (historical deadlock is fixed)
	# but measurably regresses decode 12–20% on multi-node Ray TP because each piece
	# boundary forces a cross-node sync over QSFP56 whose cost exceeds launch-overhead
	# savings. Retest only if you change away from Ray or run on a single Spark.

	COMPILATION_CONFIG='{"cudagraph_mode":"none","inductor_compile_config":{"combo_kernels":false,"benchmark_combo_kernel":false,"max_autotune":false,"max_autotune_gemm":false}}'

	# --- Speculative decoding (Agentic profile) -----------------------------------
	# ngram speculation wins on agentic / code traffic (repeated tool names, file paths,
	# JSON keys) — peak 48.34 tok/s, avg 36.44 tok/s across our 12-prompt agent set.
	# On synthetic benchmarks with low token repetition it slightly regresses decode.
	# To switch to the "Throughput-stable" profile, comment out the SPECULATIVE_CONFIG
	# line below and remove --speculative-config from the vllm serve invocation.

	SPECULATIVE_CONFIG='{"method":"ngram","num_speculative_tokens":5,"prompt_lookup_max":4,"prompt_lookup_min":2}'

	# --- vLLM serve ---------------------------------------------------------------
	exec vllm serve "$MODEL_PATH" \
	--host "$HOST" --port "$PORT" \
	--served-model-name "$SERVED_NAME" \
	--tensor-parallel-size "$TP_SIZE" \
	--distributed-executor-backend ray \
	--gpu-memory-utilization "$GPU_MEM_UTIL" \
	--max-model-len "$MAX_MODEL_LEN" \
	--max-num-seqs "$MAX_NUM_SEQS" \
	--max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" \
	--kv-cache-dtype fp8_e4m3 \
	--attention-backend flashinfer \
	--attention-config.use_trtllm_attention=0 \
	--enable-prefix-caching \
	--enable-chunked-prefill \
	--trust-remote-code \
	--enable-auto-tool-choice --tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2_append_think \
	--compilation-config "$COMPILATION_CONFIG" \
	--speculative-config "$SPECULATIVE_CONFIG"