nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 22 Packages Projects Releases Wiki Activity

model : add support for talkie-1930-13b (#22596)

Browse Source 
* initial talkie support, coherent

* reorder to follow convention

* absorb inverse rope

* stop folding scalars to improve quantization

* use broadcasting instead of duplication

* style cleanup

* add scaling support to LoraTorchTensor; use that path in conversion

* use layer_out_scale instead of embd_skip_scale
This commit is contained in:
Niklas Sheth
2026-05-26 07:57:38 +03:00
committed by GitHub
parent 1506d39e76
commit c9d98295a3
14 changed files with 263 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gguf-py/gguf/constants.py
+15
View File
@@ -505,6 +505,7 @@ class MODEL_ARCH(IntEnum):
LLAMA_EMBED = auto()
MAINCODER = auto()
KIMI_LINEAR = auto()
TALKIE = auto()
class VISION_PROJECTOR_TYPE(IntEnum):
@@ -1021,6 +1022,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.LLAMA_EMBED: "llama-embed",
MODEL_ARCH.MAINCODER: "maincoder",
MODEL_ARCH.KIMI_LINEAR: "kimi-linear",
MODEL_ARCH.TALKIE: "talkie",
}
VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -4013,6 +4015,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
MODEL_ARCH.TALKIE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.LAYER_OUT_SCALE,
],
# TODO
}
gguf-py/gguf/tensor_mapping.py
+10
View File
@@ -34,6 +34,7 @@ class TensorNameMap:
"encoder", # neobert
"model.transformer.wte", # llada
"embed_tokens", # qwen3-embedding
"model.embed", # talkie
),
# Token type embeddings
@@ -259,6 +260,7 @@ class TensorNameMap:
"model.transformer.blocks.{bid}.q_proj", # llada
"layers.{bid}.self_attn.q_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.q_proj", # nemotron-h
"model.blocks.{bid}.attn.attn_query", # talkie
),
# Attention key
@@ -279,6 +281,7 @@ class TensorNameMap:
"model.transformer.blocks.{bid}.k_proj", # llada
"layers.{bid}.self_attn.k_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.k_proj", # nemotron-h
"model.blocks.{bid}.attn.attn_key", # talkie
),
# Attention value
@@ -298,6 +301,7 @@ class TensorNameMap:
"model.transformer.blocks.{bid}.v_proj", # llada
"layers.{bid}.self_attn.v_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.v_proj", # nemotron-h
"model.blocks.{bid}.attn.attn_value", # talkie
),
# Attention output
@@ -336,6 +340,7 @@ class TensorNameMap:
"layers.{bid}.self_attn.o_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.o_proj", # nemotron-h
"model.layers.{bid}.self_attn.language_expert_dense", # cogvlm
"model.blocks.{bid}.attn.attn_resid", # talkie
),
# Attention output norm
@@ -508,6 +513,7 @@ class TensorNameMap:
"layers.{bid}.mlp.up_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.up_proj", # nemotron-h
"model.layers.{bid}.mlp.language_mlp.up_proj", # cogvlm
"model.blocks.{bid}.mlp.mlp_linear", # talkie
),
MODEL_TENSOR.FFN_UP_EXP: (
@@ -561,6 +567,7 @@ class TensorNameMap:
"model.transformer.blocks.{bid}.ff_proj", # llada
"layers.{bid}.mlp.gate_proj", # qwen3-embedding
"model.layers.{bid}.mlp.language_mlp.gate_proj", # cogvlm
"model.blocks.{bid}.mlp.mlp_gate", # talkie
),
MODEL_TENSOR.FFN_GATE_EXP: (
@@ -636,6 +643,7 @@ class TensorNameMap:
"layers.{bid}.mlp.down_proj", # qwen3-embedding
"backbone.layers.{bid}.mixer.down_proj", # nemotron-h
"model.layers.{bid}.mlp.language_mlp.down_proj", # cogvlm
"model.blocks.{bid}.mlp.mlp_resid", # talkie
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -682,6 +690,7 @@ class TensorNameMap:
"model.layers.layers.{bid}.mixer.q_norm", # plamo3
"layers.{bid}.self_attn.q_norm", # qwen3-embedding
"model.layers.{bid}.attention.query_layernorm", # apertus
"model.blocks.{bid}.attn.head_gain.head_g", # talkie
),
MODEL_TENSOR.ATTN_K_NORM: (
@@ -716,6 +725,7 @@ class TensorNameMap:
MODEL_TENSOR.LAYER_OUT_SCALE: (
"model.layers.{bid}.layer_scalar", # gemma4
"model.blocks.{bid}.embed_skip.a_g", # talkie
),
MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (