Add Vast.ai Mem0 Qwen3.5 launcher

.gitattributes

@@ -0,0 +1 @@
+*.sh text eol=lf

config/sequential_mem0_qwen35_4b.yaml

@@ -0,0 +1,38 @@
+# Mem0 preset for a local Qwen3.5-4B OpenAI-compatible endpoint.
+# The Vast.ai launcher in scripts/vastai/overrides the endpoint and served-model
+# values at runtime when the corresponding environment variables are set.
+agent_name: Structure_rag_mem0_qwen35_4b
+model: Qwen3.5-4B
+temperature: 0.7
+input_length_limit: 30000
+model_context_window: 32768
+buffer_length: 2000
+output_dir: ./results/outputs/Qwen3.5-4B-mem0
+
+# Mem0 runtime and retrieval settings.
+agent_chunk_size: 4096
+retrieve_num: 100
+mem0_add_infer: true
+
+# Local chat-completion endpoint.
+provider: openai_compatible
+api_key_env: OPENAI_API_KEY
+base_url: http://127.0.0.1:9909/v1
+base_url_env:
+azure_endpoint:
+azure_api_version:
+tokenizer_encoding: cl100k_base
+
+# Keep Qwen thinking disabled for Mem0 fact extraction and benchmark answers.
+qwen3_disable_thinking: true
+
+# Local embedding endpoint. Mem0's Qdrant collection uses the 4B model's
+# 2560-dimensional vectors.
+mem0_embedder_model: Qwen3-Embedding-4B
+embedding_api_key_env: OPENAI_API_KEY
+embedding_base_url: http://127.0.0.1:9009/v1
+embedding_base_url_env:
+
+# Optional prompt overrides. Empty values select benchmark-specific prompts.
+mem0_fact_extraction_prompt:
+mem0_memory_answer_prompt:

requirements-mem0-vast.txt

@@ -0,0 +1,21 @@
+# Minimal MemoryData dependencies for the Mem0 + local OpenAI-compatible path.
+# vLLM (and therefore torch/transformers) is installed separately by the Vast.ai
+# launcher so the model server can use a recent Qwen3.5-compatible release.
+numpy<2
+python-dotenv
+PyYAML
+tqdm
+datasets
+tiktoken
+httpx
+openai==2.30.0
+pydantic==2.12.5
+pydantic_core==2.41.5
+langchain-core==1.2.25
+qdrant-client
+posthog
+pytz
+mem0ai==1.0.10
+nltk
+rouge-score
+editdistance

scripts/vastai/README.md

@@ -0,0 +1,69 @@
+# Vast.ai: Mem0 with local Qwen3.5-4B
+
+This launcher runs the MemoryData Mem0 benchmark against two local
+OpenAI-compatible vLLM services:
+
+- chat and fact extraction: `Qwen/Qwen3.5-4B`
+- memory embeddings: `Qwen/Qwen3-Embedding-4B`
+
+The default benchmark is the MemoryAgentBench `longmemeval_s*` subset using the
+bundled 150k-context LongMemEval preset.
+
+## Instance recommendation
+
+Use a CUDA Vast.ai image with Python 3.11+ and enough disk space for both model
+checkpoints. Two 16 GB or larger GPUs are preferred, one for Qwen3.5-4B and one
+for Qwen3-Embedding-4B. A single 24 GB or larger GPU can host both services; the
+launcher automatically assigns conservative memory fractions when they share a
+GPU.
+
+The default run is a one-query smoke test:
+
+```bash
+bash scripts/vastai/run_mem0_qwen35.sh
+```
+
+Run every query in the configured five-sample LongMemEval preset:
+
+```bash
+MAX_QUERIES=0 bash scripts/vastai/run_mem0_qwen35.sh
+```
+
+Remove the preset's five-sample cap and run every available LongMemEval sample:
+
+```bash
+MAX_QUERIES=0 MAX_TEST_SAMPLES=999999 bash scripts/vastai/run_mem0_qwen35.sh
+```
+
+Run on explicit GPUs or with another dataset config:
+
+```bash
+LLM_GPU=0 \
+EMBED_GPU=1 \
+MAX_QUERIES=0 \
+DATASET_CONFIG=benchmark/membench/config/MemBench_simple.yaml \
+bash scripts/vastai/run_mem0_qwen35.sh
+```
+
+MemBench data must already exist under `datasets/MemBench/`. MemoryAgentBench
+falls back to its Hugging Face dataset when the local copy is absent.
+
+## Common overrides
+
+| Variable | Default | Purpose |
+| --- | --- | --- |
+| `MAX_QUERIES` | `1` | Smoke-test query count; use `0` for every query in the selected samples |
+| `MAX_TEST_SAMPLES` | preset value (`5` for LongMemEval) | Positive integer overriding the dataset sample cap |
+| `FORCE` | `0` | Use `1` to discard and rebuild prior result/agent state |
+| `LLM_GPU` | `0` | Physical GPU index for the chat server |
+| `EMBED_GPU` | `1` when available, otherwise `0` | Physical GPU index for embeddings |
+| `MAX_MODEL_LEN` | `32768` | Qwen3.5 vLLM context limit |
+| `LLM_MODEL` | `Qwen/Qwen3.5-4B` | Hugging Face model ID or local model directory |
+| `EMBED_MODEL` | `Qwen/Qwen3-Embedding-4B` | Embedding model ID or local directory |
+| `KEEP_SERVERS` | `0` | Use `1` to leave servers running after the benchmark |
+| `SKIP_INSTALL` | `0` | Use `1` after the virtual environment is fully prepared |
+| `HF_HOME` | `~/.cache/huggingface` | Persistent model cache location |
+
+Server logs and the generated runtime YAML are saved under
+`.runtime-cache/vastai-mem0-qwen35/`. Benchmark outputs retain the normal
+`results/outputs/` and `results/agents/` layout.