Marlin downstream clean

[robertgshaw2-redhat]

Cleaned up version of Marlin PR that can be merged into main.

Added some E2E tests, which compare the results of the exllama kernels to the marlin kernels. See tests/models/test_marlin.py for more details

• Some E2E performance data on a range of GPUs. TLDR: these kernels rock

[robertgshaw2-redhat]

To use:

from vllm import LLM, SamplingParams

model = LLM("robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin")

sampling_params = SamplingParams(max_tokens=100, temperature=0)
outputs = model.generate("Hello my name is", sampling_params=sampling_params)
outputs[0].outputs[0].text

[mgoin]

What is the "forked" argument doing here?

[robertgshaw2-redhat]

Im not sure, all the other models tests have this arg though.

I've just been running pytest tests/models/test_marlin.py

[tlrmchlsmth]

I think we'll need to include a license file for marlin somewhere

[tlrmchlsmth]

can these be const&?

[robertgshaw2-redhat]

@alexm-nm ?

[robertgshaw2-redhat]

I think we'll need to include a license file for marlin somewhere

@tlrmchlsmth

We have the following in the cuda code:

/*
 * Copyright (C) Marlin.2024 Elias Frantar (elias.frantar@ist.ac.at)
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *         http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

[robertgshaw2-redhat]

Closing in favor of #43