Kimi-Linear support (backend agnostic + MLA KV cache) (#18755)
* kimi linear model implementation * kimi linear convert_hf_to_gguf * kimi linear constants.py tensor_mapping.py * Kimi Linear ggml.h * kimi linear ggml-cpu * Kimi Linear ggml-cuda * Kimi Linear ggml.c * kimi linear src/llama * remove "const int64_t n_seq_tokens = q->ne[2];" to get rid of unused variable warning * remove type mismatch warning * read MoE params * removed some hard coded code * removed all hard code * use DeepseekV2 tokenizer * removed unnecessary internal methods called by the old set_vocab of KimiLinear * rewrite get_vocab for KimiLinear. Removed all kda_scan code * removed all traces of kda_scan * reduce OP count by 1 due to removal of kda_scan * Move KIMI_LINEAR to llm_arch_is_hybrid to enable KV cache * set n_embd_head_k/v to ensure kv cache works * don't quantize conv1d of Kimi Linear * Kimi Linear backend agnostic * removed LOG_INFO * naive chunking form implemented * fixed some comments * add Kimi-K2 specific tokens to be recognized as EOG * build_kda_autoregressive is implemented to replace build_kda_recurrent for faster inference. sync'd to b7682 * replaced Akk and Aqk with mul_mat and clamp * no clamp version * Moved Aqk computation out of the loop * fixed typo and split wkv_b into wk_b and wv_b * MLA KV cache support * fix trailing spaces * moved const llama_model & model; around to follow qwen3next format and see if it cna pass the -Wunused-private-field error * fix trailing whitespace * removed traling whitespaces in empty line + make sure indentation is multiple of 4 * try to make lint happy * remove blank lines to make lint happy * removed at least blank line containing white space * fixed flake8 complaints locally * return ggml_tensor * pair in kda_autoregressive and kda_chunking as in ngxson's Qwen3Next improvement * removed Kimi-Linear specific change that causes failure at server-windows * removed private: from kimi_linear to make build checks happy * removed unnecessary ggml_cont before ggml_reshape * created static function causal_conv1d to abtract similar code for q/k/v * merged dt_bias to SSM_DT. Do -exp(log_A) in convert_hf_to_gguf.py. * reverted to original * fixed find_hparam calls. Fixed e_score_correction_bias to use bias instead of weight. Removed all ssm_conv bias terms. * remove DT_B from constants.py. remove one comment line in llama-model.cpp * new class llm_graph_input_mem_hybrid_k to get around the new MLA change. switch the concat order of ggml_concat calls in kimi-linear.cpp to accommodate MLA changes. Removed support for exp_probs_b.weight * remove ssm_o_norm_b * remove ssm_o_norm_b * changed hparams.kda_head_dim to hparams.n_embd_head_kda. added TODO comment for class llama_graph_mem_hybrid_k * removed all ggml_cont b4 ggml_reshape_4d * Whitespace * replaced all hparams.get with find_hparams * added new names for n_experts, n_experts_used and score_func in TextModel and removed their code in KimiLinear in convert_hf_to_gguf.py. Removed unnecessary ggml_cont and GGML_ASSERT in kimi-linear.cpp * use is_mla to switch between different mem_hybrid types * fixed logical errors in convert_hf_to_gguf.py pointed out by CISC * removed if else for required parameters kv_lora_rank and qk_rope_head_dim * add back ggml_cont for Vcur * minor changes * removed extra line in llama-vocab.cpp. Added back the comment in llama-graph.cpp * f16 gguf cannot run without context length * made a mistake of adding back n_ctx parsing --------- Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
This commit is contained in:
ymcki
+222
-3
@@ -586,6 +586,10 @@ class ModelBase:
|
||||
gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
|
||||
gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
|
||||
gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
|
||||
# Kimi KDA conv weights should be F32
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_Q,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_K,
|
||||
gguf.MODEL_TENSOR.SSM_CONV1D_V,
|
||||
)
|
||||
)
|
||||
or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
|
||||
@@ -903,10 +907,10 @@ class TextModel(ModelBase):
|
||||
if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
|
||||
self.gguf_writer.add_layer_norm_eps(f_norm_eps)
|
||||
logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
|
||||
if (n_experts := self.hparams.get("num_local_experts")) is not None:
|
||||
if (n_experts := self.find_hparam(["num_local_experts", "num_experts"], optional=True)) is not None:
|
||||
self.gguf_writer.add_expert_count(n_experts)
|
||||
logger.info(f"gguf: expert count = {n_experts}")
|
||||
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
|
||||
if (n_experts_used := self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)) is not None:
|
||||
self.gguf_writer.add_expert_used_count(n_experts_used)
|
||||
logger.info(f"gguf: experts used count = {n_experts_used}")
|
||||
if (n_expert_groups := self.hparams.get("n_group")) is not None:
|
||||
@@ -916,7 +920,7 @@ class TextModel(ModelBase):
|
||||
self.gguf_writer.add_expert_group_used_count(n_group_used)
|
||||
logger.info(f"gguf: expert groups used count = {n_group_used}")
|
||||
|
||||
if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
|
||||
if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation_func"], optional=True)) is not None:
|
||||
if score_func == "sigmoid":
|
||||
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
|
||||
elif score_func == "softmax":
|
||||
@@ -5013,6 +5017,221 @@ class CodeShellModel(TextModel):
|
||||
self.gguf_writer.add_rope_scaling_factor(1.0)
|
||||
|
||||
|
||||
@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
|
||||
class KimiLinearModel(TextModel):
|
||||
"""Kimi-Linear model with hybrid MLA+KDA architecture"""
|
||||
model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
|
||||
|
||||
_experts: list[dict[str, Tensor]] | None = None
|
||||
|
||||
def set_vocab(self):
|
||||
try:
|
||||
self._set_vocab_gpt2()
|
||||
return
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
from transformers import AutoTokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
|
||||
tokpre = self.get_vocab_base_pre(tokenizer)
|
||||
|
||||
if tokpre == "kimi-k2":
|
||||
# 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)
|
||||
# override eos id in config.json with tiktoken eos id
|
||||
self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
|
||||
else:
|
||||
raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
# note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)
|
||||
self.hparams["num_key_value_heads"] = 1
|
||||
|
||||
super().set_gguf_parameters()
|
||||
self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
|
||||
|
||||
# KDA & MLA params
|
||||
# Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
|
||||
linear_attn_config = self.hparams["linear_attn_config"]
|
||||
# n_head == 0 for KDA layers, n_head > 0 for MLA layers
|
||||
# full_attention_layers list will be used to distingush layer type
|
||||
_num_kv_heads = list()
|
||||
_full_attn_layers = linear_attn_config["full_attn_layers"]
|
||||
for il in range(self.hparams["num_hidden_layers"]):
|
||||
if il + 1 in _full_attn_layers:
|
||||
_num_kv_heads.append(self.hparams["num_key_value_heads"])
|
||||
else:
|
||||
_num_kv_heads.append(0)
|
||||
assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]
|
||||
self.gguf_writer.add_head_count_kv(_num_kv_heads)
|
||||
|
||||
if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:
|
||||
self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)
|
||||
if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:
|
||||
self.gguf_writer.add_kda_head_dim(kda_head_dim)
|
||||
|
||||
# MLA params - use add_* methods that handle arch substitution
|
||||
# Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)
|
||||
if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:
|
||||
self.gguf_writer.add_q_lora_rank(q_lora_rank)
|
||||
# To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA
|
||||
kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)
|
||||
self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
|
||||
|
||||
# MLA head dimensions
|
||||
# Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim
|
||||
qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")
|
||||
# Rotation - use qk_rope_head_dim for Kimi
|
||||
qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)
|
||||
self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
|
||||
self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
|
||||
v_head_dim = self.hparams.get("v_head_dim")
|
||||
|
||||
# Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
|
||||
if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:
|
||||
self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
|
||||
elif qk_nope_head_dim is not None:
|
||||
n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
|
||||
self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
|
||||
|
||||
# n_embd_head_v_mla = v_head_dim
|
||||
if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:
|
||||
self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)
|
||||
elif v_head_dim is not None:
|
||||
self.gguf_writer.add_value_length_mla(v_head_dim)
|
||||
|
||||
# moe_intermediate_size (1024 for Kimi)
|
||||
self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
|
||||
# num_shared_experts (1 for Kimi)
|
||||
self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
|
||||
# first_k_dense_replace (1 for Kimi - first layer uses dense MLP)
|
||||
self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
|
||||
# Routed scaling factor (expert_weights_scale = 2.446 for Kimi)
|
||||
self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
|
||||
|
||||
def prepare_tensors(self):
|
||||
super().prepare_tensors()
|
||||
if self._experts is not None:
|
||||
experts = [k for d in self._experts for k in d.keys()]
|
||||
if len(experts) > 0:
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")
|
||||
|
||||
# Handle KDA conv1d weights
|
||||
# HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest
|
||||
# llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest
|
||||
# GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
|
||||
# Memory layouts match: both have conv_step (d_conv) changing fastest
|
||||
if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
|
||||
# HF shape: [d_inner, d_conv] e.g. [4096, 4]
|
||||
# Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
|
||||
if data_torch.ndim == 2:
|
||||
d_inner, d_conv = data_torch.shape
|
||||
# Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)
|
||||
data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
|
||||
logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
|
||||
elif data_torch.ndim == 3:
|
||||
# Already 3D [d_inner, 1, d_conv] from unsqueeze
|
||||
d_inner, _, d_conv = data_torch.shape
|
||||
data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
|
||||
logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, 1, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
|
||||
|
||||
# Kimi specific bias
|
||||
if name.endswith("e_score_correction_bias"):
|
||||
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
|
||||
|
||||
# Handle A_log: iHF stores as [1, 1, num_heads, 1]
|
||||
# llama.cpp expects ggml ne = [1, num_heads, 1, 1]
|
||||
# GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]
|
||||
if name.endswith(".A_log"):
|
||||
data_torch = -torch.exp(data_torch)
|
||||
if name.endswith(".dt_bias"):
|
||||
name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
|
||||
logger.info("Changed dt_bias to dt_proj.bias")
|
||||
|
||||
# process the experts separately
|
||||
if name.find("block_sparse_moe.experts") != -1:
|
||||
n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
|
||||
assert bid is not None
|
||||
|
||||
if self._experts is None:
|
||||
self._experts = [{} for _ in range(self.block_count)]
|
||||
|
||||
self._experts[bid][name] = data_torch
|
||||
|
||||
if len(self._experts[bid]) >= n_experts * 3:
|
||||
# merge the experts into a single 3d tensor
|
||||
# w1: gate, w2: down, w3: up
|
||||
for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),
|
||||
("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),
|
||||
("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:
|
||||
datas: list[Tensor] = []
|
||||
for xid in range(n_experts):
|
||||
ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
|
||||
datas.append(self._experts[bid][ename])
|
||||
del self._experts[bid][ename]
|
||||
data_torch = torch.stack(datas, dim=0)
|
||||
new_name = self.format_tensor_name(tname, bid)
|
||||
yield from super().modify_tensors(data_torch, new_name, bid)
|
||||
return
|
||||
|
||||
# note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
|
||||
if name.endswith("kv_b_proj.weight"):
|
||||
name_kb = name.replace("kv_b_proj", "k_b_proj")
|
||||
name_vb = name.replace("kv_b_proj", "v_b_proj")
|
||||
n_head_kv = self.hparams["num_key_value_heads"]
|
||||
v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)
|
||||
qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
|
||||
logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)
|
||||
assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
|
||||
kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
|
||||
k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
|
||||
k_b = k_b.transpose(1, 2)
|
||||
yield from super().modify_tensors(k_b, name_kb, bid)
|
||||
yield from super().modify_tensors(v_b, name_vb, bid)
|
||||
return
|
||||
|
||||
yield from super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
|
||||
@ModelBase.register("InternLM2ForCausalLM")
|
||||
class InternLM2Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.INTERNLM2
|
||||
|
||||
Reference in New Issue
Block a user