catid

cmake_minimum_required(VERSION 3.10)
project(MonteCarloCasino)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# Set optimization flags
set(CMAKE_CXX_FLAGS_RELEASE "-O3 -march=native -mtune=native")

# Find threads package

catid

How to quantize 70B model so it will fit on 2x4090 GPUs:

I tried EXL2, AutoAWQ, and SqueezeLLM and they all failed for different reasons (issues opened).

HQQ worked:

I rented a 4x GPU 1TB RAM ($19/hr) instance on runpod with 1024GB container and 1024GB workspace disk space.
I think you only need 2x GPU with 80GB VRAM and 512GB+ system RAM so probably overpaid.

Note you need to fill in the form to get access to the 70B Meta weights.

catid

# conda create -n dbrx python=3.10 -y && conda activate dbrx
# pip install torch transformers tiktoken flash_attn bitsandbytes

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("SinclairSchneider/dbrx-instruct-quantization-fixed", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("SinclairSchneider/dbrx-instruct-quantization-fixed", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True, load_in_4bit=True)

input_text = "What does it take to build a great LLM?"

catid

# Collaboration between Claude-3 and GPT-4 to implement https://arxiv.org/pdf/2312.02116.pdf
# This is just the GMM decoder part of the model they propose (which is the new thing).
# This one was mainly generated by GPT-4.
# The AIs provided two implementations of the idea and revised eachothers' code.
# I tested that the unit tests pass but haven't tried it in a language model yet.

import torch
import torch.nn as nn
import torch.nn.functional as F

catid

# Collaboration between Claude-3 and GPT-4 to implement https://arxiv.org/pdf/2312.02116.pdf
# This is just the GMM decoder part of the model they propose (which is the new thing).
# This one was mainly generated by Claude-3.
# The AIs provided two implementations of the idea and revised eachothers' code.
# I tested that the unit tests pass but haven't tried it in a language model yet.

import torch
import torch.nn as nn
import torch.nn.functional as F

catid

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import math

#torch.autograd.set_detect_anomaly(True)

class FeedForward(torch.nn.Module):
    def __init__(self, input_features, output_features):

catid

import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F

# This layer is dropped into your pre-trained PyTorch model where nn.Linear is used
class DoRALayer(nn.Module):
    def __init__(self, d_in, d_out, rank=4):
        super().__init__()

catid

#####################################################################
#   Auto Z-Calibration
#####################################################################

[z_calibration]
probe_nozzle_x: 175.5
probe_nozzle_y: 257
#   The X and Y coordinates (in mm) for clicking the nozzle on the
#   Z endstop.
probe_switch_x: 169.3

catid

[gcode_macro PRINT_START]
gcode:

    M117 Print Starting...

    ; Make sure we are not applying stale bed mesh or Z offset
    SET_GCODE_OFFSET Z=0
    BED_MESH_CLEAR

    ; Start heating bed

catid

[gcode_macro PRINT_START]
gcode:

    M117 Print Starting...

    ; Make sure we are not applying stale bed mesh or Z offset
    SET_GCODE_OFFSET Z=0
    BED_MESH_CLEAR

    ; Start heating bed