lmganon123/DeepSeek-R1_IK_GGUF_Q2

IQ2_KS quant of DeepSeek-R1 I made for my 192GB DDR5 + 3090/4090. Done according to:

IQ2_KS 183.004 GiB (2.339 BPW)

👈 Secret Recipe

#!/usr/bin/env bash

custom="
# First 3 dense layers (0-3) (GPU)
# Except blk.*.attn_k_b.weight is not divisible by 256 so only supports qN_0
blk\.[0-2]\.attn_k_b.*=q8_0
blk\.[0-2]\.attn_.*=iq5_ks
blk\.[0-2]\.ffn_down.*=iq5_ks
blk\.[0-2]\.ffn_(gate|up).*=iq5_ks
blk\.[0-2]\..*=iq5_ks

# All attention, norm weights, and bias tensors for MoE layers (3-60) (GPU)
# Except blk.*.attn_k_b.weight is not divisible by 256 so only supports qN_0
blk\.[3-9]\.attn_k_b.*=q8_0
blk\.[1-5][0-9]\.attn_k_b.*=q8_0
blk\.60\.attn_k_b.*=q8_0

blk\.[3-9]\.attn_.*=iq5_ks
blk\.[1-5][0-9]\.attn_.*=iq5_ks
blk\.60\.attn_.*=iq5_ks

# Shared Expert (3-60) (GPU)
blk\.[3-9]\.ffn_down_shexp\.weight=iq4_ks
blk\.[1-5][0-9]\.ffn_down_shexp\.weight=iq4_ks
blk\.60\.ffn_down_shexp\.weight=iq4_ks

blk\.[3-9]\.ffn_(gate|up)_shexp\.weight=iq4_ks
blk\.[1-5][0-9]\.ffn_(gate|up)_shexp\.weight=iq4_ks
blk\.60\.ffn_(gate|up)_shexp\.weight=iq4_ks

# Routed Experts (3-60) (CPU)
blk\.[3-9]\.ffn_down_exps\.weight=iq2_k
blk\.[1-5][0-9]\.ffn_down_exps\.weight=iq2_k
blk\.60\.ffn_down_exps\.weight=iq2_k

blk\.[3-9]\.ffn_(gate|up)_exps\.weight=iq2_ks
blk\.[1-5][0-9]\.ffn_(gate|up)_exps\.weight=iq2_ks
blk\.60\.ffn_(gate|up)_exps\.weight=iq2_ks

# Token embedding and output tensors (GPU)
token_embd\.weight=iq4_k
output\.weight=Q8_0

Prompt format

<｜begin▁of▁sentence｜>{system_prompt}<｜User｜>{prompt}<｜Assistant｜><｜end▁of▁sentence｜><｜Assistant｜>

Example run command

llama-server  --model <Path>\DeepSeek-R1-IQ2_KS-00001-of-00005.gguf -fa -rtr -mla 3 --ctx-size 40000  -ctk q8_0  -b 4092 -ub 4092 -amb 512 --n-gpu-layers 99   -ot "blk\.(3)\.ffn_.*=CUDA0" --override-tensor exps=CPU  --threads 8 --host 127.0.0.1 --port 8080

ik_llama.cpp quantizations of DeepSeek-R1

NOTE: These quants MUST be run using the llama.cpp fork, ik_llama.cpp

Credits to @ubergarm for his DeepSeek quant recipes for which these quants were based on.

Credits to @ggfhez for his bf16 upload.

Credits to @bartowski for his imatrix