CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add LLaDA 8b Diffusion model (#14771)

Browse Source 
* Add support for Llada-8b: diffusion model

* Add README

* Fix README and convert_hf_to_gguf

* convert_hf_to_gguf.py: address review comments

* Make everything in a single example

* Remove model-specific sampling

* Remove unused argmax

* Remove braced initializers, improve README.md a bit

* Add diffusion specific gguf params in set_vocab, remove setting rope_theta and rms_norm_eps

* Remove adding the mask token

* Move add_add_bos_token to set_vocab

* use add_bool in gguf_writer.py
This commit is contained in:
Aman Gupta
2025-07-31 19:49:09 +08:00
committed by GitHub
parent 11490b3672
commit 8a4a856277
12 changed files with 931 additions and 385 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
gguf-py/gguf/constants.py
View File

@@ -279,6 +279,9 @@ class Keys:
class Projector:
STACK_FACTOR = "clip.audio.projector.stack_factor"
class Diffusion:
SHIFT_LOGITS = "diffusion.shift_logits"
#
# recommended mapping of model tensor names for storage in gguf
#
@@ -377,6 +380,7 @@ class MODEL_ARCH(IntEnum):
LFM2 = auto()
DREAM = auto()
SMALLTHINKER = auto()
LLADA = auto()
class VISION_PROJECTOR_TYPE(IntEnum):
@@ -697,6 +701,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.LFM2: "lfm2",
MODEL_ARCH.DREAM: "dream",
MODEL_ARCH.SMALLTHINKER: "smallthinker",
MODEL_ARCH.LLADA: "llada",
}
VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1318,6 +1323,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.LLADA: [
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.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.QWEN2VL: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,

5
gguf-py/gguf/gguf_writer.py
View File

@@ -1047,6 +1047,11 @@ class GGUFWriter:
def add_audio_stack_factor(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
# diffusion models
def add_diffusion_shift_logits(self, value: bool) -> None:
self.add_bool(Keys.Diffusion.SHIFT_LOGITS, value)
def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
pack_prefix = ''
if not skip_pack_prefix:

12
gguf-py/gguf/tensor_mapping.py
View File

@@ -32,6 +32,7 @@ class TensorNameMap:
"model.word_embeddings", # bailingmoe
"language_model.model.embed_tokens", # llama4
"encoder", # neobert
"model.transformer.wte", # llada
),
# Token type embeddings
@@ -71,6 +72,7 @@ class TensorNameMap:
"head", # rwkv
"head.out", # wavtokenizer
"lm_head", # llama4
"model.transformer.ff_out", # llada
),
# Output norm
@@ -94,6 +96,7 @@ class TensorNameMap:
"model.ln_out", # rwkv7
"backbone.final_layer_norm", # wavtokenizer
"model.norm", # llama4
"model.transformer.ln_f", # llada
),
# Rope frequencies
@@ -139,6 +142,7 @@ class TensorNameMap:
"model.layers.{bid}.input_layernorm", # llama4
"transformer_encoder.{bid}.attention_norm", # neobert
"model.layers.{bid}.operator_norm", # lfm2
"model.transformer.blocks.{bid}.attn_norm", # llada
),
# Attention norm 2
@@ -183,6 +187,7 @@ class TensorNameMap:
"transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
"transformer.h.{bid}.attn.attention.q_proj", # exaone
"model.layers.{bid}.self_attn.q_proj", # llama4
"model.transformer.blocks.{bid}.q_proj", # llada
),
# Attention key
@@ -199,6 +204,7 @@ class TensorNameMap:
"transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
"transformer.h.{bid}.attn.attention.k_proj", # exaone
"model.layers.{bid}.self_attn.k_proj", # llama4
"model.transformer.blocks.{bid}.k_proj", # llada
),
# Attention value
@@ -214,6 +220,7 @@ class TensorNameMap:
"transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok
"transformer.h.{bid}.attn.attention.v_proj", # exaone
"model.layers.{bid}.self_attn.v_proj", # llama4
"model.transformer.blocks.{bid}.v_proj", # llada
),
# Attention output
@@ -246,6 +253,7 @@ class TensorNameMap:
"transformer.h.{bid}.attn.attention.out_proj", # exaone
"model.layers.{bid}.self_attn.o_proj", # llama4
"transformer_encoder.{bid}.wo", # neobert
"model.transformer.blocks.{bid}.attn_out", # llada
),
# Attention output norm
@@ -291,6 +299,7 @@ class TensorNameMap:
"model.layers.{bid}.post_attention_layernorm", # llama4
"transformer_encoder.{bid}.ffn_norm", # neobert
"model.layers.layers.{bid}.pre_mlp_norm", # plamo2
"model.transformer.blocks.{bid}.ff_norm", # llada
),
# Post feed-forward norm
@@ -364,6 +373,7 @@ class TensorNameMap:
"model.layers.{bid}.feed_forward.up_proj", # llama4 jamba granite-hybrid
"transformer_encoder.{bid}.ffn.w12", # neobert
"model.layers.{bid}.block_sparse_moe.up", # smallthinker
"model.transformer.blocks.{bid}.up_proj", # llada
),
MODEL_TENSOR.FFN_UP_EXP: (
@@ -405,6 +415,7 @@ class TensorNameMap:
"transformer.h.{bid}.mlp.c_fc_0", # exaone
"model.layers.{bid}.feed_forward.gate_proj", # llama4 jamba granite-hybrid
"model.layers.{bid}.block_sparse_moe.gate", # smallthinker
"model.transformer.blocks.{bid}.ff_proj", # llada
),
MODEL_TENSOR.FFN_GATE_EXP: (
@@ -454,6 +465,7 @@ class TensorNameMap:
"model.layers.{bid}.feed_forward.down_proj", # llama4 jamba granite-hybrid
"transformer_encoder.{bid}.ffn.w3", # neobert
"model.layers.{bid}.block_sparse_moe.down", # smallthinker
"model.transformer.blocks.{bid}.ff_out", # llada
),
MODEL_TENSOR.FFN_DOWN_EXP: (