compute1/Agents-A1-GPTQ-INT4-Sym

Agents-A1-GPTQ-INT4-Sym

Symmetric INT4 GPTQ quantization of InternScience/Agents-A1, a 35B Mixture-of-Experts agentic model with 256 experts and hybrid Mamba/attention architecture.

Why this quant?

The community AWQ-INT4 release (cyankiwi/Agents-A1-AWQ-INT4) uses asymmetric INT4 quantization via the compressed-tensors format. SGLang's compressed-tensors MoE path only supports symmetric weight-only quantization for the Marlin MoE kernel — the _is_wNa16_group_channel gate in get_moe_scheme() requires symmetric=True. Asymmetric quants fail this check and fall through to the W8A8 detection path, which crashes with AttributeError: 'NoneType' object has no attribute 'num_bits' because input_activations is None in weight-only quantization.

This model was re-quantized with symmetric GPTQ to be Marlin-compatible, matching the format used by the working palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4 model (which uses sym=True, group_size=128).

Quantization Details

Ignored layers (kept in BF16/FP16)

These layers are excluded from quantization for quality and Marlin kernel compatibility:

• —.*attn.* — attention projections (hybrid linear/full attention)
• —.*mlp\.gate$ — MoE router (must NOT be quantized for Marlin MoE kernel)
• —.*shared_expert.* — shared expert MLP
• —.*mtp.* — multi-token prediction head
• —.*visual.* — vision encoder (not needed for text-only serving)
• —lm_head — output projection
• —model.language_model.embed_tokens — token embeddings

Result: 30,720 of 31,181 layers quantized to INT4. All 256 MoE expert projections per layer (gateproj, upproj, down_proj × 256 experts × 40 layers) are quantized.

Model Architecture

Deployment

SGLang

python -m sglang.launch_server \
  --model-path compute1/Agents-A1-GPTQ-INT4-Sym \
  --served-model-name agents-a1 \
  --trust-remote-code \
  --quantization moe_wna16 \
  --tensor-parallel-size 2 \
  --context-length 262144 \
  --mem-fraction-static 0.87 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --enable-hierarchical-cache \
  --enable-mixed-chunk \
  --dtype float16 \
  --host 0.0.0.0 \
  --port 8000

vLLM

vllm serve compute1/Agents-A1-GPTQ-INT4-Sym \
  --served-model-name agents-a1 \
  --trust-remote-code \
  --tensor-parallel-size 2 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Recommended Sampling Parameters

From the original model card:

• —temperature: 0.85
• —top_p: 0.95
• —top_k: 20
• —min_p: 0.0
• —presence_penalty: 1.1
• —repetition_penalty: 1.0

VRAM Requirements

Measured on 2× RTX 3090 (24 GB each) with TP=2, --quantization moe_wna16, --mem-fraction-static 0.87, --dtype float16:

The model fits on 2× RTX 3090 (24 GB each). For additional KV cache headroom, --kv-cache-dtype fp8_e5m2 can be used to double KV token capacity.

Quantization was performed on 4× RTX 3090 with pipeline parallelism, using ~68 GB peak RAM and ~16 GB peak VRAM on the primary GPU.

Acknowledgements

• —InternScience for the original Agents-A1 model
• —Intel AutoRound team for the quantization toolkit
• —Model card adapted from the original model card

License

Apache-2.0, inherited from InternScience/Agents-A1.