Kimi-K2-Instruct-DRAFT-0.6B-v3.0

A 0.6B parameter draft (speculative decoding) model for use with Kimi-K2-Instruct.

See Kimi-K2-Instruct-DRAFT-0.6B-v3.0-GGUF for the models in gguf format for use with llama.cpp.

---

Extending the context above 32k

The current config.json is set for context length up to 32k tokens. Add the "rope_scaling" section to config.json to enable YaRN, eg:

To extend the context to 64k:

  "max_position_embeddings": 65536,
  ...
  "rope_scaling": {
    "factor": 2.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  },

To extend the context to 128k:

  "max_position_embeddings": 131072,
  ...
  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  },

NOTE: Because llama.cpp uses "static-YaRN" the scaling factor remains constant regardless of input length! Only add the rope_scaling configuration when processing long contexts is required...

---

How this model was created

1. The initial model was created from Qwen2.5-0.5B-Instruct using transplant-vocab:

python ./transplant_vocab.py \
    ./Qwen2.5-0.5B-Instruct \
    ./Kimi-K2-Instruct \
    ./Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED \
    --trust-remote-code \
    --override "[BOS]" "<|endoftext|>" \
    --override "[EOS]" "<|im_end|>" \
    --override "<|im_end|>" "<|im_end|>" \
    --override "<|im_user|>" "<|im_start|>user" \
    --override "<|im_assistant|>" "<|im_start|>assistant" \
    --override "<|start_header_id|>" "<|im_start|>" \
    --override "<|end_header_id|>" "<|im_end|>" \
    --override "[EOT]" "<|endoftext|>" \
    --override "<|im_system|>" "<|im_start|>system" \
    --override "<|tool_calls_section_begin|>" "<tool_call>" \
    --override "<|tool_calls_section_end|>" "</tool_call>" \
    --override "<|tool_call_begin|>" "<tool_call>" \
    --override "<|tool_call_argument_begin|>" "<tool_call>" \
    --override "<|tool_call_end|>" "</tool_call>" \
    --override "<|im_middle|>" "\\n" \
    --override "[UNK]" "<|endoftext|>" \
    --override "[PAD]" "<|endoftext|>"

Loading config from 'Qwen2.5-0.5B-Instruct'... Done.
Loading config from 'Kimi-K2-Instruct'... Done.
Loading tokenizer from 'Qwen2.5-0.5B-Instruct'... Done.
Loading tokenizer from 'Kimi-K2-Instruct'... Done.
Loading model from 'Qwen2.5-0.5B-Instruct'... Done.

Input model configuration:
- Target vocabulary size    : 163840 (used = 163840, unused = 0)
- Donor vocabulary size     : 151936
- Donor num layers          : 24 (tied embeddings = True)
- Donor hidden size         : 896
- Donor attention heads     : 14
- Donor intermediate size   : 4864 (ratio = 1:5.4)
- Donor total parameters    : 494032768 (0.49B)
-- Embedding parameters     : 136134656 (0.14B)
-- Non-embedding parameters : 357898112 (0.36B)

Processing 3 automatic token overrides:
✔ 'bos_token_id' : 163584 '[BOS]' → [151643] '<|endoftext|>'
✔ 'eos_token_id' : 163585 '[EOS]' → [151645] '<|im_end|>'
✔ 'pad_token_id' : 163839 '[PAD]' → [151643] '<|endoftext|>'

Processing 17 manual token overrides:
✔ 163584 : '[BOS]' → [151643] '<|endoftext|>'
✔ 163585 : '[EOS]' → [151645] '<|im_end|>'
✔ 163586 : '<|im_end|>' → [151645] '<|im_end|>'
✔ 163587 : '<|im_user|>' → [151644, 872] '<|im_start|>user'
✔ 163588 : '<|im_assistant|>' → [151644, 77091] '<|im_start|>assistant'
✔ 163590 : '<|start_header_id|>' → [151644] '<|im_start|>'
✔ 163591 : '<|end_header_id|>' → [151645] '<|im_end|>'
✔ 163593 : '[EOT]' → [151643] '<|endoftext|>'
✔ 163594 : '<|im_system|>' → [151644, 8948] '<|im_start|>system'
✔ 163595 : '<|tool_calls_section_begin|>' → [151657] '<tool_call>'
✔ 163596 : '<|tool_calls_section_end|>' → [151658] '</tool_call>'
✔ 163597 : '<|tool_call_begin|>' → [151657] '<tool_call>'
✔ 163598 : '<|tool_call_argument_begin|>' → [151657] '<tool_call>'
✔ 163599 : '<|tool_call_end|>' → [151658] '</tool_call>'
✔ 163601 : '<|im_middle|>' → [198] '\n'
✔ 163838 : '[UNK]' → [151643] '<|endoftext|>'
✔ 163839 : '[PAD]' → [151643] '<|endoftext|>'

NOTE: Using an "untied" copy of 'embed_tokens.weight' as new 'lm_head.weight' tensor...

Transplanting tokens: 100%|████████████████████████████████████████████████████████████| 163840/163840 [01:08<00:00, 2406.47token/s]

Transplant mappings:
- 1 to 1  : 95449 (58%)
- 2 to 1  : 61938 (38%)
- 3 to 1  : 4995 (3%)
- 4 to 1  : 980 (0.6%)
- 5 to 1  : 147 (0.09%)
- 6 to 1  : 52 (0.032%)
- 7 to 1  : 15 (0.0092%)
- 8 to 1  : 17 (0.01%)
- 9 to 1  : 2 (0.0012%)
- 10 to 1 : 5 (0.0031%)
- 11 to 1 : 1 (0.00061%)
- 13 to 1 : 239 (0.15%)

Head initialized with:
- Copies : 95449 (58%)
- Means  : 68391 (42%)
- Zeros  : 0 (0%)

Output model configuration:
- Output vocabulary size    : 163840
- Output num layers         : 24 (tied embeddings = False)
- Output hidden size        : 896
- Output attention heads    : 14
- Output intermediate size  : 4864 (ratio = 1:5.4)
- Output total parameters   : 651499392 (0.65B)
-- Embedding parameters     : 293601280 (0.29B)
-- Non-embedding parameters : 357898112 (0.36B)

Saving model and tokenizer to 'Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED' folder
[2025-08-07 15:47:15,620] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)

Patching 'torch_dtype' in 'Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED/config.json' based on actual saved tensors
- Updated 'torch_dtype' to 'bfloat16' based on actual tensor dtype

Operation completed successfully (ignore any 'segmentation fault' that follows!!!)

NOTE: Due to the non-standard tokenizer, this needs the --trust-remote-code option.

NOTE: I had to manually delete "pad_token_id": 163839 from config.json to get it to match the tokeniser when used in llama.cpp as a draft model.

2. The following datasets were used to create a fine-tuning dataset of ~2.3B tokens:

• agentlans/common-crawl-sample
• bigcode/the-stack-smol-xl
• rombodawg/Everything_Instruct (NOTE: output field only)

formatted just between [EOS] tags.

3. The model was then trained using qlora-pipe-lite for 1 epoch with a batch size of 60 and a sequence length of 32k (~2M tokens per step):

# ==============================
# MODEL AND OUTPUT CONFIGURATION
# ==============================

model_dir = 'models/Kimi-K2-Instruct-DRAFT-0.6B-UNTRAINED'
output_dir = 'finetuned'

# ===========================
# TRAINING TYPE CONFIGURATION
# ===========================

full_fine_tune = true

# =======================
# OPTIMIZER CONFIGURATION
# =======================

lr = 5e-5

# ======================
# TRAINING CONFIGURATION
# ======================

sequence_len = 32768

gradient_accumulation_steps = 10  # 10×6 = batch size 60, 10×6×32768 = ~2M tokens per step

# =====================
# DATASET CONFIGURATION
# =====================

[[datasets]]
dataset_path = 'datasets/common-crawl-sample/*.json'
drop_tails = true

[[datasets]]
dataset_path = 'datasets/the-stack-smol-xl/*.jsonl'
drop_tails = true

[[datasets]]
dataset_path = 'datasets/rombodawg-Everything-Instruct/*.json'
drop_tails = true

NOTE: Due to the non-standard tokenizer, this needs the --trust-remote-code option passing on the deepspeed call to train.py.

I used six RTX A6000 GPUs over three nodes and hence the 60 batch size (6 x 10 gradient accumulation steps = 60):

4. Fixing the TikToken / SentencePiece tokenizer mismatch in llama.cpp

I had to temporarily hack this change into convert_hf_to_gguf.py:

@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration")
class Qwen2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.QWEN2

    #def set_vocab(self):
    #    try:
    #        self._set_vocab_sentencepiece()
    #    except FileNotFoundError:
    #        self._set_vocab_gpt2()
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
        tokpre = self.get_vocab_base_pre(tokenizer)
    
        # Build merges list using the approach similar to HunYuanMoE
        merges = []
        vocab = {}
        mergeable_ranks = tokenizer.model._mergeable_ranks
        for token, rank in mergeable_ranks.items():
            vocab[QwenModel.token_bytes_to_string(token)] = rank
            if len(token) == 1:
                continue
            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
            if len(merged) == 2:
                merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))

        # Build token list
        vocab_size = self.hparams["vocab_size"]
        special_tokens = tokenizer.special_tokens
        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
        tokens: list[str] = []
        toktypes: list[int] = []
    
        for i in range(vocab_size):
            if i not in reverse_vocab:
                tokens.append(f"[PAD{i}]")
                toktypes.append(gguf.TokenType.UNUSED)
            else:
                token = reverse_vocab[i]
                tokens.append(token)
                if i in special_tokens.values():
                    toktypes.append(gguf.TokenType.CONTROL)
                else:
                    toktypes.append(gguf.TokenType.NORMAL)
    
        self.gguf_writer.add_tokenizer_model("gpt2")
        self.gguf_writer.add_tokenizer_pre(tokpre)
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)
        self.gguf_writer.add_token_merges(merges)

        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
        special_vocab.add_to_gguf(self.gguf_writer)

This then let me run:

~/llama.cpp/convert_hf_to_gguf.py --outtype auto --outfile Kimi-K2-Instruct-DRAFT-0.6B-BF16.gguf Kimi-K2-Instruct-DRAFT-0.6B

and then it quantized OK:

~/llama.cpp/build/bin/llama-quantize Kimi-K2-Instruct-DRAFT-0.6B-BF16.gguf Kimi-K2-Instruct-DRAFT-0.6B-Q4_0.gguf Q4_0 44