---

base_model: ibm-granite/granite-3.0-3b-a800m-instruct frameworks:

• PyTorch library_name: openmind license: apache-2.0 pipeline_tag: text-generation tags:
• language
• granite-3.0 quantized_by: bartowski inference: false model-index:
• name: granite-3.0-2b-instruct results:
  • task: type: text-generation dataset: name: IFEval type: instruction-following metrics:
    • type: pass@1 value: 42.49 name: pass@1
    • type: pass@1 value: 7.02 name: pass@1
  • task: type: text-generation dataset: name: AGI-Eval type: human-exams metrics:
    • type: pass@1 value: 25.7 name: pass@1
    • type: pass@1 value: 50.16 name: pass@1
    • type: pass@1 value: 20.51 name: pass@1
  • task: type: text-generation dataset: name: OBQA type: commonsense metrics:
    • type: pass@1 value: 40.8 name: pass@1
    • type: pass@1 value: 59.95 name: pass@1
    • type: pass@1 value: 71.86 name: pass@1
    • type: pass@1 value: 67.01 name: pass@1
    • type: pass@1 value: 48.0 name: pass@1
  • task: type: text-generation dataset: name: BoolQ type: reading-comprehension metrics:
    • type: pass@1 value: 78.65 name: pass@1
    • type: pass@1 value: 6.71 name: pass@1
  • task: type: text-generation dataset: name: ARC-C type: reasoning metrics:
    • type: pass@1 value: 50.94 name: pass@1
    • type: pass@1 value: 26.85 name: pass@1
    • type: pass@1 value: 37.7 name: pass@1
  • task: type: text-generation dataset: name: HumanEvalSynthesis type: code metrics:
    • type: pass@1 value: 39.63 name: pass@1
    • type: pass@1 value: 40.85 name: pass@1
    • type: pass@1 value: 35.98 name: pass@1
    • type: pass@1 value: 27.4 name: pass@1
  • task: type: text-generation dataset: name: GSM8K type: math metrics:
    • type: pass@1 value: 47.54 name: pass@1
    • type: pass@1 value: 19.86 name: pass@1
  • task: type: text-generation dataset: name: PAWS-X (7 langs) type: multilingual metrics:
    • type: pass@1 value: 50.23 name: pass@1
    • type: pass@1 value: 28.87 name: pass@1

---

Llamacpp imatrix Quantizations of granite-3.0-3b-a800m-instruct

Using llama.cpp release b3930 for quantization.

Original model: https://huggingface.co/ibm-granite/granite-3.0-3b-a800m-instruct

All quants made using imatrix option with dataset from here

Run them in LM Studio

Prompt format

<|start_of_role|>system<|end_of_role|>{system_prompt}<|end_of_text|>
<|start_of_role|>user<|end_of_role|>{prompt}<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>

Use in openmind

Note: Currently trans doesn't support loading granitemoe related models, writing them manually can load them, but there may be problems with the tokenizer.

import os
import time
import argparse
import torch
from torch import nn
import numpy as np
import logging

from transformers.integrations import GGUF_TENSOR_MAPPING, GGUF_CONFIG_MAPPING
from transformers.modeling_gguf_pytorch_utils import GGUF_TO_TRANSFORMERS_MAPPING, GGUF_SUPPORTED_ARCHITECTURES
from transformers.integrations.ggml import GGUF_TO_FAST_CONVERTERS, GGUFGPTConverter

model_type = "granitemoe"

GGUF_TENSOR_MAPPING.update(
    {
        model_type: {
            "token_embd": "model.embed_tokens",
            "blk": "model.layers",
            "ffn_gate_exps": "block_sparse_moe.gate_linear",
            "ffn_up_exps": "block_sparse_moe.input_linear",
            "ffn_down_exps": "block_sparse_moe.output_linear",
            "ffn_gate_inp": "block_sparse_moe.router.layer",
            "ffn_norm": "post_attention_layernorm",
            "attn_norm": "input_layernorm",
            "attn_q": "self_attn.q_proj",
            "attn_v": "self_attn.v_proj",
            "attn_k": "self_attn.k_proj",
            "attn_output": "self_attn.o_proj",
            "output.weight": "lm_head.weight",
            "output_norm": "model.norm",
        }
    }
)

GGUF_CONFIG_MAPPING.update(
    {
        model_type: {
            "block_count": "num_hidden_layers",
            "context_length": "max_position_embeddings",
            "embedding_length": "hidden_size",
            "feed_forward_length": "intermediate_size",
            "attention.head_count": "num_attention_heads",
            "attention.head_count_kv": "num_key_value_heads",
            "rope.freq_base": "rope_theta",
            "attention.layer_norm_rms_epsilon": "rms_norm_eps",
            "expert_count": "num_local_experts",
            "expert_used_count": "num_experts_per_tok",
            "vocab_size": "vocab_size",
            "rope.dimension_count": None,
        },
    }
)

GGUF_TO_TRANSFORMERS_MAPPING.update(
    {
        "config": GGUF_CONFIG_MAPPING,
        "tensors": GGUF_TENSOR_MAPPING,
    }
)

GGUF_SUPPORTED_ARCHITECTURES.append(model_type)

GGUF_TO_FAST_CONVERTERS.update(
    {
        model_type: GGUFGPTConverter,
    }
)

# 修改 transformers 包中的 GraniteMoeMoE 类
def modify_granitemoe():
    import transformers.models.granitemoe.modeling_granitemoe as granitemoe

    class ModifiedGraniteMoeMoE(nn.Module):
        def __init__(self, config: granitemoe.GraniteMoeConfig):
            super(ModifiedGraniteMoeMoE, self).__init__()

            self.input_size = config.hidden_size
            self.hidden_size = config.intermediate_size
            self.activation = granitemoe.ACT2FN[config.hidden_act]
            
            self.gate_linear = granitemoe.GraniteMoeParallelExperts(config.num_local_experts, self.input_size, self.hidden_size)
            self.input_linear = granitemoe.GraniteMoeParallelExperts(config.num_local_experts, self.input_size, self.hidden_size)
            self.output_linear = granitemoe.GraniteMoeParallelExperts(config.num_local_experts, self.hidden_size, self.input_size)
            self.router = granitemoe.GraniteMoeTopKGating(
                input_size=self.input_size,
                num_experts=config.num_local_experts,
                top_k=config.num_experts_per_tok,
            )

        def forward(self, layer_input):
            bsz, length, emb_size = layer_input.size()
            layer_input = layer_input.reshape(-1, emb_size)
            _, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)

            expert_inputs = layer_input[batch_index]
            hidden_states = self.activation(self.gate_linear(expert_inputs, expert_size)) * self.input_linear(expert_inputs, expert_size)
            expert_outputs = self.output_linear(hidden_states, expert_size)

            expert_outputs = expert_outputs * batch_gates[:, None]

            zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
            layer_output = zeros.index_add(0, batch_index, expert_outputs)
            layer_output = layer_output.view(bsz, length, self.input_size)
            return layer_output, router_logits

    granitemoe.GraniteMoeMoE = ModifiedGraniteMoeMoE

# 在导入 transformers 之前调用修改函数
modify_granitemoe()
# ------------------------------------------------

def parse_args():
    parser = argparse.ArgumentParser(description="NPU Inference for Text Generation Model")
    parser.add_argument(
        "--model_name_or_path",
        "-m",
        type=str,
        help="Path to model",
        default=".",
    )
    parser.add_argument(
        "--inference_mode",
        "-i",
        type=str,
        help="Inference mode",
        default="gguf",
    )
    parser.add_argument(
        "--gguf_file",
        "-g",
        type=str,
        help="Path to GGUF file",
        default="granite-3.0-3b-a800m-instruct-Q4_0.gguf",
    )
    parser.add_argument(
        "--debug",
        action="store_true",
        help="Debug mode",
    )
    return parser.parse_args()

def set_logging(model_name):
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
        handlers=[
            logging.FileHandler(f"{model_name}_inference.log"),
            logging.StreamHandler(),
        ],
    )

args = parse_args()
set_logging(os.path.basename(args.model_name_or_path))

if args.debug:
    logging.info("Debug mode enabled, using transformers package from source.")
    from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, is_torch_npu_available
else:
    logging.info("Debug mode disabled, using openmind package.")
    from openmind import AutoTokenizer, AutoModelForCausalLM, pipeline, is_torch_npu_available

def load_model_from_gguf(model_path: str, device_map="auto"):
    gguf_filename = args.gguf_file
    tokenizer = AutoTokenizer.from_pretrained(model_path, gguf_file=gguf_filename, tokenizer_type="gpt2")
    tokenizer.pad_token = tokenizer.eos_token
    model = AutoModelForCausalLM.from_pretrained(model_path, gguf_file=gguf_filename, device_map=device_map)
    return tokenizer, model

def load_model_from_local(model_path: str, device_map="auto"):
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device_map)
    return tokenizer, model

def load_model_from_pipeline(model_path: str, device_map="auto", task="text-generation"):
    pipeline_pt = pipeline(
        task=task,
        model=model_path,
        device_map=device_map,
        framework="pt",
        truncation=True,
    )
    return pipeline_pt.tokenizer, pipeline_pt

def load_model(mode: str, *args, **kwargs):
    if mode == "gguf":
        return load_model_from_gguf(*args, **kwargs)
    elif mode == "model":
        return load_model_from_local(*args, **kwargs)
    elif mode == "pipeline":
        return load_model_from_pipeline(*args, **kwargs)
    else:
        raise ValueError(f"load_model Unknown mode: {mode}")
    

def generate_text_form_model(tokenizer, model, prompt, max_new_tokens=50):
    inputs = tokenizer(prompt, return_tensors="pt", padding=True).to(model.device)
    output = model.generate(
        input_ids=inputs['input_ids'], 
        attention_mask=inputs['attention_mask'],
        max_new_tokens=max_new_tokens,
    )
    return tokenizer.decode(output[0], skip_special_tokens=True)

def generate_text_from_pipeline(tokenizer, pipeline, prompt, max_new_tokens=50):
    results = pipeline(
        prompt,
        max_new_tokens=max_new_tokens,
    )
    return results[0]["generated_text"]

def generate_text(mode: str, *args, **kwargs):
    if mode == "model" or mode == "gguf":
        return generate_text_form_model(*args, **kwargs)
    elif mode == "pipeline":
        return generate_text_from_pipeline(*args, **kwargs)
    else:
        raise ValueError(f"generate_text Unknown mode: {mode}")

def apply_chat_template(tokenizer, tokenize=False):
    if tokenizer.chat_template is None:
        print("Chat template is not defined, use default template.")
        tokenizer.chat_template = "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
    chat = [
        {
            "role": "system",
            "content": "You are a helpful assistant who always responds in a friendly manner",
        },
        {
            "role": "user",
            "content": "Why does the ocean appear blue?",
        },
    ]
    chat_input = tokenizer.apply_chat_template(chat, tokenize=tokenize)
    return chat_input

def main():
    model_path = args.model_name_or_path
    abs_model_path = os.path.abspath(model_path)
    model_name = os.path.basename(abs_model_path)
    logging.info(f"测试模型: {model_name}")
    logging.info(f"模型路径: {model_path}")
    logging.info(f"绝对路径: {abs_model_path}")
    inference_mode = args.inference_mode
    logging.info(f"推理模式: {inference_mode}")
    
    # 确保使用 NPU 设备
    device_map = "auto" if is_torch_npu_available() else "cpu"
    logging.info(f"NPU {'available' if device_map == 'auto' else 'not available'}, use device_map='{device_map}'.")

    # 加载模型
    tokenizer, task_pipeline = load_model(mode=inference_mode, model_path=model_path, device_map=device_map)
    prompt = apply_chat_template(tokenizer, tokenize=False)

    # 推理性能测试
    inference_times = []
    num_runs = 10

    logging.info(f"\n=== NPU {model_name} 性能测试 ===")

    for i in range(num_runs):
        input_text = prompt

        start_time = time.time()
        
        results = generate_text(inference_mode, tokenizer, task_pipeline, input_text)
        # torch.npu.synchronize()

        inference_time = time.time() - start_time
        inference_times.append(inference_time)

        if i == 0:
            logging.info(f"输入文本: {input_text}")
            logging.info("生成结果：")
            logging.info(f"  {results}")

    avg_time = np.mean(inference_times)
    std_time = np.std(inference_times)

    logging.info("\n性能分析:")
    logging.info(f"NPU平均推理时间: {avg_time:.4f} 秒")
    logging.info(f"NPU推理时间标准差: {std_time:.4f} 秒")
    logging.info(f"推理时间列表: {inference_times}")


if __name__ == "__main__":
    main()

Download a file (not the whole branch) from below:

Filename	Quant type	File Size	Split	Description
granite-3.0-3b-a800m-instruct-f16.gguf	f16	6.75GB	false	Full F16 weights.
granite-3.0-3b-a800m-instruct-Q8_0.gguf	Q8_0	3.59GB	false	Extremely high quality, generally unneeded but max available quant.
granite-3.0-3b-a800m-instruct-Q6_K_L.gguf	Q6_K_L	2.81GB	false	Uses Q8_0 for embed and output weights. Very high quality, near perfect, recommended.
granite-3.0-3b-a800m-instruct-Q6_K.gguf	Q6_K	2.78GB	false	Very high quality, near perfect, recommended.
granite-3.0-3b-a800m-instruct-Q5_K_L.gguf	Q5_K_L	2.45GB	false	Uses Q8_0 for embed and output weights. High quality, recommended.
granite-3.0-3b-a800m-instruct-Q5_K_M.gguf	Q5_K_M	2.41GB	false	High quality, recommended.
granite-3.0-3b-a800m-instruct-Q5_K_S.gguf	Q5_K_S	2.34GB	false	High quality, recommended.
granite-3.0-3b-a800m-instruct-Q4_K_L.gguf	Q4_K_L	2.12GB	false	Uses Q8_0 for embed and output weights. Good quality, recommended.
granite-3.0-3b-a800m-instruct-Q4_K_M.gguf	Q4_K_M	2.06GB	false	Good quality, default size for must use cases, recommended.
granite-3.0-3b-a800m-instruct-Q4_K_S.gguf	Q4_K_S	1.94GB	false	Slightly lower quality with more space savings, recommended.
granite-3.0-3b-a800m-instruct-Q4_0_8_8.gguf	Q4_0_8_8	1.93GB	false	Optimized for ARM inference. Requires 'sve' support (see link below). Don't use on Mac or Windows.
granite-3.0-3b-a800m-instruct-Q4_0_4_8.gguf	Q4_0_4_8	1.93GB	false	Optimized for ARM inference. Requires 'i8mm' support (see link below). Don't use on Mac or Windows.
granite-3.0-3b-a800m-instruct-Q4_0_4_4.gguf	Q4_0_4_4	1.93GB	false	Optimized for ARM inference. Should work well on all ARM chips, pick this if you're unsure. Don't use on Mac or Windows.
granite-3.0-3b-a800m-instruct-Q4_0.gguf	Q4_0	1.93GB	false	Legacy format, generally not worth using over similarly sized formats
granite-3.0-3b-a800m-instruct-Q3_K_XL.gguf	Q3_K_XL	1.84GB	false	Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability.
granite-3.0-3b-a800m-instruct-IQ4_XS.gguf	IQ4_XS	1.82GB	false	Decent quality, smaller than Q4_K_S with similar performance, recommended.
granite-3.0-3b-a800m-instruct-Q3_K_L.gguf	Q3_K_L	1.77GB	false	Lower quality but usable, good for low RAM availability.
granite-3.0-3b-a800m-instruct-Q3_K_M.gguf	Q3_K_M	1.64GB	false	Low quality.
granite-3.0-3b-a800m-instruct-IQ3_M.gguf	IQ3_M	1.52GB	false	Medium-low quality, new method with decent performance comparable to Q3_K_M.
granite-3.0-3b-a800m-instruct-Q3_K_S.gguf	Q3_K_S	1.49GB	false	Low quality, not recommended.
granite-3.0-3b-a800m-instruct-IQ3_XS.gguf	IQ3_XS	1.41GB	false	Lower quality, new method with decent performance, slightly better than Q3_K_S.
granite-3.0-3b-a800m-instruct-Q2_K_L.gguf	Q2_K_L	1.34GB	false	Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable.
granite-3.0-3b-a800m-instruct-Q2_K.gguf	Q2_K	1.27GB	false	Very low quality but surprisingly usable.

Embed/output weights

Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.

Some say that this improves the quality, others don't notice any difference. If you use these models PLEASE COMMENT with your findings. I would like feedback that these are actually used and useful so I don't keep uploading quants no one is using.

Thanks!

Downloading using huggingface-cli

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download bartowski/granite-3.0-3b-a800m-instruct-GGUF --include "granite-3.0-3b-a800m-instruct-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download bartowski/granite-3.0-3b-a800m-instruct-GGUF --include "granite-3.0-3b-a800m-instruct-Q8_0/*" --local-dir ./

You can either specify a new local-dir (granite-3.0-3b-a800m-instruct-Q8_0) or download them all in place (./)

Q4_0_X_X

These are NOT for Metal (Apple) offloading, only ARM chips.

If you're using an ARM chip, the Q4_0_X_X quants will have a substantial speedup. Check out Q4_0_4_4 speed comparisons on the original pull request

To check which one would work best for your ARM chip, you can check AArch64 SoC features (thanks EloyOn!).

Which file should I choose?

A great write up with charts showing various performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.

The I-quants are not compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.

Credits

Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset

Thank you ZeroWw for the inspiration to experiment with embed/output

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski