Skip to content

Commit

Permalink
llama : add grok-1 support (ggerganov#6204)
Browse files Browse the repository at this point in the history 
* Add support for Grok model architecture

* Revert convert-hf-to-gguf to default options

* Fixed f_norm_rms_eps bug

* Fix whitespaces

* llama : fix grok rope type

* llama : minor

---------

Co-authored-by: Georgi Gerganov <[email protected]>
  • Loading branch information
@arki05 @ggerganov @hodlen
2 people authored and hodlen committed Apr 1, 2024
1 parent 0d215d8 commit a7a2d75
Show file tree
Hide file tree
Showing 4 changed files with 384 additions and 20 deletions.
26 changes: 26 additions & 0 deletions convert-hf-to-gguf.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -93,31 +93,42 @@ def set_gguf_parameters(self):

if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
self.gguf_writer.add_context_length(n_ctx)
print(f"gguf: context length = {n_ctx}")

n_embd = self.find_hparam(["hidden_size", "n_embd"])
self.gguf_writer.add_embedding_length(n_embd)
print(f"gguf: embedding length = {n_embd}")

if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
print(f"gguf: feed forward length = {n_ff}")

n_head = self.find_hparam(["num_attention_heads", "n_head"])
self.gguf_writer.add_head_count(n_head)
print(f"gguf: head count = {n_head}")

if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
self.gguf_writer.add_head_count_kv(n_head_kv)
print(f"gguf: key-value head count = {n_head_kv}")

if (rope_theta := self.hparams.get("rope_theta")) is not None:
self.gguf_writer.add_rope_freq_base(rope_theta)
print(f"gguf: rope theta = {rope_theta}")
if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
print(f"gguf: rms norm epsilon = {f_rms_eps}")
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)
print(f"gguf: layer norm epsilon = {f_norm_eps}")
if (n_experts := self.hparams.get("num_local_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
print(f"gguf: expert count = {n_experts}")
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
print(f"gguf: experts used count = {n_experts_used}")

self.gguf_writer.add_file_type(self.ftype)
print(f"gguf: file type = {self.ftype}")

def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
Expand Down Expand Up @@ -1051,6 +1062,21 @@ def set_vocab(self):
self._set_vocab_sentencepiece()


@Model.register("GrokForCausalLM")
class GrokModel(Model):
model_arch = gguf.MODEL_ARCH.GROK

def set_vocab(self):
self._set_vocab_sentencepiece()

def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)

def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_name("Grok")


@Model.register("MiniCPMForCausalLM")
class MiniCPMModel(Model):
model_arch = gguf.MODEL_ARCH.MINICPM
Expand Down
24 changes: 24 additions & 0 deletions gguf-py/gguf/constants.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -100,6 +100,7 @@ class MODEL_ARCH(IntEnum):
LLAMA = auto()
FALCON = auto()
BAICHUAN = auto()
GROK = auto()
GPT2 = auto()
GPTJ = auto()
GPTNEOX = auto()
Expand Down Expand Up @@ -167,6 +168,7 @@ class MODEL_TENSOR(IntEnum):
MODEL_ARCH.LLAMA: "llama",
MODEL_ARCH.FALCON: "falcon",
MODEL_ARCH.BAICHUAN: "baichuan",
MODEL_ARCH.GROK: "grok",
MODEL_ARCH.GPT2: "gpt2",
MODEL_ARCH.GPTJ: "gptj",
MODEL_ARCH.GPTNEOX: "gptneox",
Expand Down Expand Up @@ -251,6 +253,28 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.GROK: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.ATTN_OUT_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.LAYER_OUT_NORM,
],
MODEL_ARCH.GPTNEOX: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
Expand Down
55 changes: 35 additions & 20 deletions gguf-py/gguf/tensor_mapping.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@ class TensorNameMap:
"model.embedding", # mamba-qbert
"backbone.embedding", # mamba
"backbone.embeddings", # mamba-hf
"transformer.in_out_embed", # Grok
),

# Token type embeddings
Expand Down Expand Up @@ -66,6 +67,7 @@ class TensorNameMap:
"lm_head.ln", # phi2
"model.norm_f", # mamba-qbert
"backbone.norm_f", # mamba
"transformer.rms_norm", # Grok
),

# Rope frequencies
Expand Down Expand Up @@ -93,6 +95,7 @@ class TensorNameMap:
"model.layers.{bid}.attention_norm", # internlm2
"model.layers.{bid}.norm", # mamba-qbert
"backbone.layers.{bid}.norm", # mamba
"transformer.decoder_layer.{bid}.rms_norm", # Grok
),

# Attention norm 2
Expand All @@ -116,32 +119,35 @@ class TensorNameMap:

# Attention query
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq" # internlm2
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.query" # Grok
),

# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
"model.layers.{bid}.attention.wk" # internlm2
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
"model.layers.{bid}.attention.wk", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok
),

# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
"model.layers.{bid}.attention.wv" # internlm2
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
"model.layers.{bid}.attention.wv", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.value" # Grok
),

# Attention output
Expand All @@ -162,12 +168,14 @@ class TensorNameMap:
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
"model.layers.{bid}.attention.wo", # internlm2
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
"transformer.decoder_layer.{bid}.multi_head_attention.linear"# Grok
),

# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
),

# Rotary embeddings
Expand All @@ -190,11 +198,13 @@ class TensorNameMap:
"model.layers.{bid}.ln2", # yi
"h.{bid}.ln_2", # gpt2
"model.layers.{bid}.ffn_norm", # internlm2
"transformer.decoder_layer.{bid}.rms_norm_2", # Grok
),

MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
"transformer.decoder_layer.{bid}.router" # Grok
),

# Feed-forward up
Expand Down Expand Up @@ -223,6 +233,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w3", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
"transformer.decoder_layer.{bid}.moe.{xid}.linear_v", # Grok
),

# AWQ-activation gate
Expand All @@ -243,6 +254,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w1", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
"transformer.decoder_layer.{bid}.moe.{xid}.linear" # Grok
),

# Feed-forward down
Expand Down Expand Up @@ -270,6 +282,8 @@ class TensorNameMap:
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w2", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
"transformer.decoder_layer.{bid}.moe.{xid}.linear_1", # Grok

),

MODEL_TENSOR.ATTN_Q_NORM: (
Expand All @@ -287,8 +301,9 @@ class TensorNameMap:
),

MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
),

MODEL_TENSOR.SSM_IN: (
Expand Down
Loading

0 comments on commit a7a2d75

Please sign in to comment.