agnitra 0.2.0

Inference optimizer for decoder-only LLMs. One-line drop-in for HuggingFace models.

Agnitra

The inference optimizer for decoder-only LLMs. One line, no retraining, faster than torch.compile on the architectures you actually run in production.

import agnitra

result = agnitra.optimize(model, input_shape=(1, 512))
fast_model = result.optimized_model

That's it. No retraining. No graph rewrites by hand. No serving stack to adopt.

Supported architectures

Agnitra is intentionally narrow. The wedge is decoder-only LLMs — Llama-class models that account for ~80% of LLM inference spend in production. Every fine-tune of every supported architecture inherits the optimization decisions of its base model via architecture fingerprinting, so "13 architectures supported" effectively means "the ~100K decoder-LM fine-tunes on HuggingFace."

Architecture	model_type	Reference model	Status
Llama 1/2/3	llama	meta-llama/Meta-Llama-3-8B-Instruct	✅ tuned specialist
Mistral	mistral	mistralai/Mistral-7B-Instruct-v0.3	✅ tuned specialist
Mixtral	mixtral	mistralai/Mixtral-8x7B-Instruct-v0.1	✅ tuned specialist
Qwen 2 / 2.5	qwen2	Qwen/Qwen2.5-7B-Instruct	✅ tuned specialist
Qwen 2 MoE	qwen2_moe	Qwen/Qwen2.5-MoE	✅ tuned specialist
Gemma 1 / 2	gemma / gemma2	google/gemma-2-9b-it	✅ tuned specialist
Phi / Phi-3	phi / phi3	microsoft/Phi-3-mini-4k-instruct	🟡 generic decoder-LM
DeepSeek V2	deepseek_v2	deepseek-ai/DeepSeek-V2-Lite	🟡 generic decoder-LM
OLMo, Yi, Falcon	olmo / yi / falcon	allenai/OLMo-7B	🟡 generic decoder-LM
Encoder transformers (BERT, RoBERTa, ViT)	—	—	❌ pass-through
Image generation (SDXL, FLUX)	—	—	❌ pass-through (ring 2)
Speech (Whisper)	—	—	❌ pass-through (ring 3)

When a model is outside the ring-1 set, agnitra.optimize returns the input model unchanged with result.notes["passthrough"] = True and the detected architecture string. Honest scoping is a feature — a silent 5% no-op speedup destroys customer trust faster than honest refusal.

LoRA fine-tunes are supported via peft.merge_and_unload() first; hot-swappable adapters are not yet supported.

Roadmap rings

• Ring 1 (now): decoder-only LLMs. Llama, Mistral, Qwen, Gemma, Phi, DeepSeek, etc.
• Ring 2 (planned): image generation. SDXL, SD3, FLUX. Different optimization landscape (UNet attention, classifier-free guidance batching, VAE decode).
• Ring 3 (planned): speech. Whisper, Wav2Vec2.
• Out of scope: encoder transformers, multimodal pipelines, image classification, training-time optimization, multi-GPU sharding.

Status

Beta. The optimizer works end-to-end on real models. Public benchmark numbers vs. torch.compile / vLLM / TensorRT-LLM are pending the first H100 run — see benchmarks/llama3_h100/RESULTS.md. Until those numbers are published, treat any "2x faster" claim as unverified.

Install

pip install agnitra

Optional extras:

pip install "agnitra[openai]"   # LLM-guided kernel suggestions via OpenAI
pip install "agnitra[rl]"       # PPO-guided search (Stable-Baselines3)
pip install "agnitra[nvml]"     # GPU telemetry via pynvml

Quickstart

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import agnitra

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct", torch_dtype=torch.float16
).cuda()

result = agnitra.optimize(model, input_shape=(1, 512), enable_rl=False)
fast = result.optimized_model

# Use `fast` everywhere you used `model` before.

A complete runnable script lives at examples/quickstart.py.

Want a real speedup? Quantize.

The TF32 + SDPA + torch.compile defaults match what HuggingFace does today; the optimization that actually beats transformers baseline on Llama-class models is INT8 weight-only quantization:

result = agnitra.optimize(
    model,
    input_shape=(1, 512),
    enable_rl=False,
    quantize="int8_weight",   # ← the lever
)

Expected impact on Llama-3-8B: ~1.3–1.7× throughput on memory-bound decode, ~2× memory reduction, near-zero quality loss. Requires torchao (pip install "agnitra[quantize]").

Integrations

HuggingFace transformers

Replace AutoModelForCausalLM with AgnitraModel. Everything else stays identical:

from agnitra.integrations.huggingface import AgnitraModel

model = AgnitraModel.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct",
    torch_dtype=torch.float16,
    agnitra_kwargs={"input_shape": (1, 512)},
).cuda()
# Use `model` like a normal transformers model — tokenizer, .generate(), logits.

Pass any transformers.AutoModelFor... class via model_class= for non-CausalLM workloads. See examples/quickstart_hf.py.

For an existing transformers.pipeline():

from agnitra.integrations.huggingface import optimize_pipeline

pipe = transformers.pipeline("text-generation", model="meta-llama/Meta-Llama-3-8B-Instruct")
optimize_pipeline(pipe, agnitra_kwargs={"input_shape": (1, 512)})

accelerate

For users who go through accelerate.Accelerator, run Agnitra after prepare():

from accelerate import Accelerator
from agnitra.integrations.accelerate_helpers import optimize_after_prepare

accelerator = Accelerator()
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
model = optimize_after_prepare(model, input_shape=(1, 512))

What Agnitra does

1. Profiles the model on real input shapes via torch.profiler + NVML telemetry.
2. Suggests kernel tuning parameters using either an LLM (OpenAI / Ollama) or a deterministic policy when no LLM is available.
3. Applies safe-by-default optimizations: TF32 matmul, FlashAttention / SDPA, torch.compile with the right mode, optional fused Triton kernels for matmul / layer-norm.
4. Verifies the optimized model produces the same outputs as the baseline.
5. Returns the patched nn.Module plus a structured report (RuntimeOptimizationResult).

What Agnitra is not

Honest scope, so you don't waste a day:

• Not a serving runtime. It does not implement paged KV cache, continuous batching, or speculative decoding. Pair Agnitra with vLLM / TGI / your own serving stack.
• Limited quantization (W8A16 only). Agnitra supports INT8 weight-only quantization via quantize="int8_weight", delegating to torchao. This is the optimization that beats plain HuggingFace + torch.compile (HF doesn't quantize by default). INT4 / activation quantization / AWQ / GPTQ are out of scope; if you have a model already quantized via those, Agnitra will optimize it but won't re-quantize.
• Not a trainer. Inference only. Training-time optimization is out of scope.
• Not a multi-GPU sharder. Single-GPU optimization. Tensor parallelism is a separate problem.

Benchmarks

Reproducible benchmarks live in benchmarks/. The headline suite is benchmarks/llama3_h100/ — a one-command repro comparing Agnitra to HuggingFace transformers, torch.compile, vLLM, and TensorRT-LLM on Llama-3-8B at batch sizes 1, 8, 32.

Five access paths are documented (Docker, host venv, Modal, Lambda Labs / RunPod SSH, GitHub Actions self-hosted). The cheapest is Modal:

pip install modal && modal token new
HF_TOKEN=hf_xxx modal run benchmarks/llama3_h100/modal_runner.py

RESULTS.md is regenerated by the benchmark CI workflow on every release tag and gates merges on a >5% throughput regression vs. the previous baseline.

LangChain

Agents call the LLM hundreds of times per task. A 1.5x speedup on the model becomes a 1.5x reduction in the agent's wall-clock time.

from langchain_huggingface import HuggingFacePipeline
from agnitra.integrations.langchain import optimize_llm

llm = HuggingFacePipeline.from_model_id("meta-llama/Meta-Llama-3-8B-Instruct", task="text-generation")
optimize_llm(llm, agnitra_kwargs={"input_shape": (1, 512), "quantize": "int8_weight"})
# Use `llm` exactly as before — all chains/agents downstream get the speedup.

See examples/quickstart_langchain.py.

LlamaIndex

Same pattern for RAG and agent flows.

from llama_index.llms.huggingface import HuggingFaceLLM
from agnitra.integrations.llama_index import optimize_llm

llm = HuggingFaceLLM(model_name="meta-llama/Meta-Llama-3-8B-Instruct", ...)
optimize_llm(llm, agnitra_kwargs={"input_shape": (1, 512), "quantize": "int8_weight"})

See examples/quickstart_llama_index.py.

CLI

agnitra optimize --model my_model.pt --input-shape 1,3,224,224 --output optimized.pt
agnitra optimize-dir --models-dir /var/agnitra/fleet  # batch-optimize a fine-tune farm
agnitra doctor                    # health check: torch / CUDA / NVML / Ollama / license
agnitra heartbeat --interval 30   # background re-optimization daemon

Fine-tune farms (agnitra optimize-dir)

If you run 50+ fine-tuned variants of the same base model in production (one per customer, one per use-case), Agnitra's architecture-fingerprint cache means the second model with the same architecture inherits the first's optimization decisions. Optimizing 50 Llama-3-8B fine-tunes takes nearly the same wall time as optimizing one.

agnitra optimize-dir \
  --models-dir /var/agnitra/fleet \
  --quantize int8_weight \
  --input-shape 1,512

The directory layout is the standard HuggingFace shape — one subdirectory per model, each containing config.json plus weights.

API server (optional)

If you want to call the optimizer remotely (CI workers, hosted inference):

agnitra-api    # binds to 127.0.0.1:8080 by default; AGNITRA_API_HOST overrides

The server exposes POST /optimize, GET /jobs/{id}, GET /health, and WebSocket /ws/jobs/{id} for live status. By default it listens on localhost; set AGNITRA_ALLOW_PUBLIC_BIND=1 if you intentionally bind to a public interface.

Configuration

Environment variable	Purpose
OPENAI_API_KEY	Enables the LLM-guided suggestion path.
AGNITRA_OLLAMA_URL	Local LLM backend (default http://localhost:11434).
AGNITRA_API_HOST / AGNITRA_API_PORT	API server bind interface.
AGNITRA_LICENSE_PATH	Path to a license file when using enterprise features.
AGNITRA_NOTIFY_WEBHOOK_URL	POST optimization results to Slack / Discord / Telegram.

Repository layout

agnitra/
  sdk.py                    public optimize() entry point
  cli.py                    Click CLI
  _sdk/                     low-level optimizer (FX, kernels, RL)
  core/
    optimizer/              LLM- and RL-guided optimization
    runtime/                runtime patching, telemetry, fingerprinting, cache
    kernel/                 Triton kernel generation
    metering/, billing/     usage events and Stripe integration
    licensing/              license validation
    notifications/          webhook notifiers
  api/                      Starlette REST API server
benchmarks/                 reproducible benchmark suites
examples/                   small focused examples
tests/                      pytest suite

Contributing

Bug reports, benchmark PRs, and "you're handicapping vLLM" issues are all welcome — the benchmark suite is meant to be adversarially reviewed. Open an issue or PR.

License

Apache 2.0.