Skip to content

Add T5Gemma support #14940 #15123

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
+550 −2
Open

Add T5Gemma support #14940 #15123

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
93b3228
arch : add T5Gemma encoder-decoder architecture support (#14940)
baonudesifeizhai Aug 6, 2025
f5144c1
fix: add type safety checks for T5Gemma model initialization
baonudesifeizhai Aug 6, 2025
f892c18
arch : add T5Gemma encoder-decoder architecture support with improvem…
baonudesifeizhai Aug 6, 2025
062de3a
fix: resolve type annotation issues in T5Gemma model initialization
baonudesifeizhai Aug 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
297 changes: 297 additions & 0 deletions convert_hf_to_gguf.py
View file
Open in desktop

Large diffs are not rendered by default.

53 changes: 53 additions & 0 deletions gguf-py/gguf/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -362,6 +362,7 @@ class MODEL_ARCH(IntEnum):
BITNET = auto()
T5 = auto()
T5ENCODER = auto()
T5GEMMA = auto() # T5Gemma architecture
JAIS = auto()
NEMOTRON = auto()
EXAONE = auto()
Expand Down Expand Up @@ -528,6 +529,12 @@ class MODEL_TENSOR(IntEnum):
DEC_FFN_DOWN = auto()
DEC_FFN_UP = auto()
DEC_OUTPUT_NORM = auto()
# T5GEMMA specific post layer normalization tensors
DEC_POST_SELF_ATTN_NORM = auto()
DEC_POST_CROSS_ATTN_NORM = auto()
DEC_POST_FFN_NORM = auto()
ENC_POST_SELF_ATTN_NORM = auto()
ENC_POST_FFN_NORM = auto()
ENC_ATTN_NORM = auto()
ENC_ATTN_Q = auto()
ENC_ATTN_K = auto()
Expand Down Expand Up @@ -693,6 +700,7 @@ class MODEL_TENSOR(IntEnum):
MODEL_ARCH.BITNET: "bitnet",
MODEL_ARCH.T5: "t5",
MODEL_ARCH.T5ENCODER: "t5encoder",
MODEL_ARCH.T5GEMMA: "t5gemma", # T5Gemma architecture
MODEL_ARCH.JAIS: "jais",
MODEL_ARCH.NEMOTRON: "nemotron",
MODEL_ARCH.EXAONE: "exaone",
Expand Down Expand Up @@ -860,6 +868,12 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.DEC_FFN_DOWN: "dec.blk.{bid}.ffn_down",
MODEL_TENSOR.DEC_FFN_UP: "dec.blk.{bid}.ffn_up",
MODEL_TENSOR.DEC_OUTPUT_NORM: "dec.output_norm",
# T5GEMMA specific post layer normalization tensors
MODEL_TENSOR.DEC_POST_SELF_ATTN_NORM: "dec.blk.{bid}.post_self_attn_norm",
MODEL_TENSOR.DEC_POST_CROSS_ATTN_NORM: "dec.blk.{bid}.post_cross_attn_norm",
MODEL_TENSOR.DEC_POST_FFN_NORM: "dec.blk.{bid}.post_ffn_norm",
MODEL_TENSOR.ENC_POST_SELF_ATTN_NORM: "enc.blk.{bid}.post_self_attn_norm",
MODEL_TENSOR.ENC_POST_FFN_NORM: "enc.blk.{bid}.post_ffn_norm",
MODEL_TENSOR.ENC_ATTN_NORM: "enc.blk.{bid}.attn_norm",
MODEL_TENSOR.ENC_ATTN_Q: "enc.blk.{bid}.attn_q",
MODEL_TENSOR.ENC_ATTN_K: "enc.blk.{bid}.attn_k",
Expand Down Expand Up @@ -2238,6 +2252,45 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.ENC_FFN_UP,
MODEL_TENSOR.ENC_OUTPUT_NORM,
],
MODEL_ARCH.T5GEMMA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.DEC_ATTN_NORM,
MODEL_TENSOR.DEC_ATTN_Q,
MODEL_TENSOR.DEC_ATTN_K,
MODEL_TENSOR.DEC_ATTN_V,
MODEL_TENSOR.DEC_ATTN_OUT,
MODEL_TENSOR.DEC_ATTN_REL_B,
MODEL_TENSOR.DEC_CROSS_ATTN_NORM,
MODEL_TENSOR.DEC_CROSS_ATTN_Q,
MODEL_TENSOR.DEC_CROSS_ATTN_K,
MODEL_TENSOR.DEC_CROSS_ATTN_V,
MODEL_TENSOR.DEC_CROSS_ATTN_OUT,
MODEL_TENSOR.DEC_CROSS_ATTN_REL_B,
MODEL_TENSOR.DEC_FFN_NORM,
MODEL_TENSOR.DEC_FFN_GATE,
MODEL_TENSOR.DEC_FFN_DOWN,
MODEL_TENSOR.DEC_FFN_UP,
MODEL_TENSOR.DEC_OUTPUT_NORM,
MODEL_TENSOR.ENC_ATTN_NORM,
MODEL_TENSOR.ENC_ATTN_Q,
MODEL_TENSOR.ENC_ATTN_K,
MODEL_TENSOR.ENC_ATTN_V,
MODEL_TENSOR.ENC_ATTN_OUT,
MODEL_TENSOR.ENC_ATTN_REL_B,
MODEL_TENSOR.ENC_FFN_NORM,
MODEL_TENSOR.ENC_FFN_GATE,
MODEL_TENSOR.ENC_FFN_DOWN,
MODEL_TENSOR.ENC_FFN_UP,
MODEL_TENSOR.ENC_OUTPUT_NORM,
# T5GEMMA specific post layer normalization tensors
MODEL_TENSOR.DEC_POST_SELF_ATTN_NORM,
MODEL_TENSOR.DEC_POST_CROSS_ATTN_NORM,
MODEL_TENSOR.DEC_POST_FFN_NORM,
MODEL_TENSOR.ENC_POST_SELF_ATTN_NORM,
MODEL_TENSOR.ENC_POST_FFN_NORM,
],
MODEL_ARCH.JAIS: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
Expand Down
44 changes: 44 additions & 0 deletions gguf-py/gguf/tensor_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,8 @@ class TensorNameMap:
"embedding.word_embeddings", # chatglm
"transformer.token_embeddings", # openelm
"shared", # t5
"model.decoder.embed_tokens", # t5gemma
"model.encoder.embed_tokens", # t5gemma
"rwkv.embeddings", # rwkv6
"model.embeddings", # rwkv7
"model.word_embeddings", # bailingmoe
Expand Down Expand Up @@ -887,22 +889,27 @@ class TensorNameMap:

MODEL_TENSOR.DEC_ATTN_NORM: (
"decoder.block.{bid}.layer.0.layer_norm", # t5
"model.decoder.layers.{bid}.pre_self_attn_layernorm", # t5gemma
),

MODEL_TENSOR.DEC_ATTN_Q: (
"decoder.block.{bid}.layer.0.SelfAttention.q", # t5
"model.decoder.layers.{bid}.self_attn.q_proj", # t5gemma
),

MODEL_TENSOR.DEC_ATTN_K: (
"decoder.block.{bid}.layer.0.SelfAttention.k", # t5
"model.decoder.layers.{bid}.self_attn.k_proj", # t5gemma
),

MODEL_TENSOR.DEC_ATTN_V: (
"decoder.block.{bid}.layer.0.SelfAttention.v", # t5
"model.decoder.layers.{bid}.self_attn.v_proj", # t5gemma
),

MODEL_TENSOR.DEC_ATTN_OUT: (
"decoder.block.{bid}.layer.0.SelfAttention.o", # t5
"model.decoder.layers.{bid}.self_attn.o_proj", # t5gemma
),

MODEL_TENSOR.DEC_ATTN_REL_B: (
Expand All @@ -911,22 +918,27 @@ class TensorNameMap:

MODEL_TENSOR.DEC_CROSS_ATTN_NORM: (
"decoder.block.{bid}.layer.1.layer_norm", # t5
"model.decoder.layers.{bid}.pre_cross_attn_layernorm", # t5gemma
),

MODEL_TENSOR.DEC_CROSS_ATTN_Q: (
"decoder.block.{bid}.layer.1.EncDecAttention.q", # t5
"model.decoder.layers.{bid}.cross_attn.q_proj", # t5gemma
),

MODEL_TENSOR.DEC_CROSS_ATTN_K: (
"decoder.block.{bid}.layer.1.EncDecAttention.k", # t5
"model.decoder.layers.{bid}.cross_attn.k_proj", # t5gemma
),

MODEL_TENSOR.DEC_CROSS_ATTN_V: (
"decoder.block.{bid}.layer.1.EncDecAttention.v", # t5
"model.decoder.layers.{bid}.cross_attn.v_proj", # t5gemma
),

MODEL_TENSOR.DEC_CROSS_ATTN_OUT: (
"decoder.block.{bid}.layer.1.EncDecAttention.o", # t5
"model.decoder.layers.{bid}.cross_attn.o_proj", # t5gemma
),

MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: (
Expand All @@ -935,43 +947,70 @@ class TensorNameMap:

MODEL_TENSOR.DEC_FFN_NORM: (
"decoder.block.{bid}.layer.2.layer_norm", # t5
"model.decoder.layers.{bid}.pre_feedforward_layernorm", # t5gemma
),

MODEL_TENSOR.DEC_FFN_GATE: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi_0", # flan-t5
"model.decoder.layers.{bid}.mlp.gate_proj", # t5gemma
),

MODEL_TENSOR.DEC_FFN_UP: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi", # t5
"decoder.block.{bid}.layer.2.DenseReluDense.wi_1", # flan-t5
"model.decoder.layers.{bid}.mlp.up_proj", # t5gemma
),

MODEL_TENSOR.DEC_FFN_DOWN: (
"decoder.block.{bid}.layer.2.DenseReluDense.wo", # t5
"model.decoder.layers.{bid}.mlp.down_proj", # t5gemma
),

MODEL_TENSOR.DEC_OUTPUT_NORM: (
"decoder.final_layer_norm", # t5
"model.decoder.norm", # t5gemma
),

# T5GEMMA specific post layer normalization tensors
MODEL_TENSOR.DEC_POST_SELF_ATTN_NORM: (
"model.decoder.layers.{bid}.post_self_attn_layernorm", # t5gemma
),
MODEL_TENSOR.DEC_POST_CROSS_ATTN_NORM: (
"model.decoder.layers.{bid}.post_cross_attn_layernorm", # t5gemma
),
MODEL_TENSOR.DEC_POST_FFN_NORM: (
"model.decoder.layers.{bid}.post_feedforward_layernorm", # t5gemma
),
MODEL_TENSOR.ENC_POST_SELF_ATTN_NORM: (
"model.encoder.layers.{bid}.post_self_attn_layernorm", # t5gemma
),
MODEL_TENSOR.ENC_POST_FFN_NORM: (
"model.encoder.layers.{bid}.post_feedforward_layernorm", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_NORM: (
"encoder.block.{bid}.layer.0.layer_norm", # t5
"model.encoder.layers.{bid}.pre_self_attn_layernorm", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_Q: (
"encoder.block.{bid}.layer.0.SelfAttention.q", # t5
"model.encoder.layers.{bid}.self_attn.q_proj", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_K: (
"encoder.block.{bid}.layer.0.SelfAttention.k", # t5
"model.encoder.layers.{bid}.self_attn.k_proj", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_V: (
"encoder.block.{bid}.layer.0.SelfAttention.v", # t5
"model.encoder.layers.{bid}.self_attn.v_proj", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_OUT: (
"encoder.block.{bid}.layer.0.SelfAttention.o", # t5
"model.encoder.layers.{bid}.self_attn.o_proj", # t5gemma
),

MODEL_TENSOR.ENC_ATTN_REL_B: (
Expand All @@ -980,25 +1019,30 @@ class TensorNameMap:

MODEL_TENSOR.ENC_FFN_NORM: (
"encoder.block.{bid}.layer.1.layer_norm", # t5
"model.encoder.layers.{bid}.pre_feedforward_layernorm", # t5gemma
),

MODEL_TENSOR.ENC_FFN_GATE: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi_0", # flan-t5
"model.encoder.layers.{bid}.mlp.gate_proj", # t5gemma
),

MODEL_TENSOR.ENC_FFN_UP: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi", # t5
"encoder.block.{bid}.layer.1.DenseReluDense.wi_1", # flan-t5
"model.encoder.layers.{bid}.mlp.up_proj", # t5gemma
),

MODEL_TENSOR.ENC_FFN_DOWN: (
"encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
"model.encoder.layers.{bid}.mlp.down_proj", # t5gemma
),

############################################################################
# TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
MODEL_TENSOR.ENC_OUTPUT_NORM: (
"encoder.final_layer_norm", # t5
"model.encoder.norm", # t5gemma
"layer_norm", # neobert
),

Expand Down
46 changes: 46 additions & 0 deletions src/llama-arch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_BITNET, "bitnet" },
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_T5ENCODER, "t5encoder" },
{ LLM_ARCH_T5GEMMA, "t5gemma" },
{ LLM_ARCH_JAIS, "jais" },
{ LLM_ARCH_NEMOTRON, "nemotron" },
{ LLM_ARCH_EXAONE, "exaone" },
Expand Down Expand Up @@ -1499,6 +1500,46 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" },
},
},
{
LLM_ARCH_T5GEMMA,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_DEC_OUTPUT_NORM, "dec.output_norm" },
{ LLM_TENSOR_DEC_ATTN_NORM, "dec.blk.%d.attn_norm" },
{ LLM_TENSOR_DEC_ATTN_Q, "dec.blk.%d.attn_q" },
{ LLM_TENSOR_DEC_ATTN_K, "dec.blk.%d.attn_k" },
{ LLM_TENSOR_DEC_ATTN_V, "dec.blk.%d.attn_v" },
{ LLM_TENSOR_DEC_ATTN_OUT, "dec.blk.%d.attn_o" },
{ LLM_TENSOR_DEC_ATTN_REL_B, "dec.blk.%d.attn_rel_b" },
{ LLM_TENSOR_DEC_CROSS_ATTN_NORM, "dec.blk.%d.cross_attn_norm" },
{ LLM_TENSOR_DEC_CROSS_ATTN_Q, "dec.blk.%d.cross_attn_q" },
{ LLM_TENSOR_DEC_CROSS_ATTN_K, "dec.blk.%d.cross_attn_k" },
{ LLM_TENSOR_DEC_CROSS_ATTN_V, "dec.blk.%d.cross_attn_v" },
{ LLM_TENSOR_DEC_CROSS_ATTN_OUT, "dec.blk.%d.cross_attn_o" },
{ LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "dec.blk.%d.cross_attn_rel_b" },
{ LLM_TENSOR_DEC_FFN_NORM, "dec.blk.%d.ffn_norm" },
{ LLM_TENSOR_DEC_FFN_GATE, "dec.blk.%d.ffn_gate" },
{ LLM_TENSOR_DEC_FFN_DOWN, "dec.blk.%d.ffn_down" },
{ LLM_TENSOR_DEC_FFN_UP, "dec.blk.%d.ffn_up" },
{ LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
{ LLM_TENSOR_ENC_ATTN_NORM, "enc.blk.%d.attn_norm" },
{ LLM_TENSOR_ENC_ATTN_Q, "enc.blk.%d.attn_q" },
{ LLM_TENSOR_ENC_ATTN_K, "enc.blk.%d.attn_k" },
{ LLM_TENSOR_ENC_ATTN_V, "enc.blk.%d.attn_v" },
{ LLM_TENSOR_ENC_ATTN_OUT, "enc.blk.%d.attn_o" },
{ LLM_TENSOR_ENC_ATTN_REL_B, "enc.blk.%d.attn_rel_b" },
{ LLM_TENSOR_ENC_FFN_NORM, "enc.blk.%d.ffn_norm" },
{ LLM_TENSOR_ENC_FFN_GATE, "enc.blk.%d.ffn_gate" },
{ LLM_TENSOR_ENC_FFN_DOWN, "enc.blk.%d.ffn_down" },
{ LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" },
{ LLM_TENSOR_DEC_POST_SELF_ATTN_NORM, "dec.blk.%d.post_self_attn_norm" },
{ LLM_TENSOR_DEC_POST_CROSS_ATTN_NORM, "dec.blk.%d.post_cross_attn_norm" },
{ LLM_TENSOR_DEC_POST_FFN_NORM, "dec.blk.%d.post_ffn_norm" },
{ LLM_TENSOR_ENC_POST_SELF_ATTN_NORM, "enc.blk.%d.post_self_attn_norm" },
{ LLM_TENSOR_ENC_POST_FFN_NORM, "enc.blk.%d.post_ffn_norm" },
},
},
{
LLM_ARCH_JAIS,
{
Expand Down Expand Up @@ -2196,6 +2237,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_ENC_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_DEC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
{LLM_TENSOR_ENC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
{LLM_TENSOR_DEC_POST_SELF_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_DEC_POST_CROSS_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_DEC_POST_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ENC_POST_SELF_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ENC_POST_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
{LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
Expand Down
7 changes: 7 additions & 0 deletions src/llama-arch.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,6 +70,7 @@ enum llm_arch {
LLM_ARCH_BITNET,
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
LLM_ARCH_T5GEMMA,
LLM_ARCH_JAIS,
LLM_ARCH_NEMOTRON,
LLM_ARCH_EXAONE,
Expand Down Expand Up @@ -381,6 +382,12 @@ enum llm_tensor {
LLM_TENSOR_DEC_FFN_DOWN,
LLM_TENSOR_DEC_FFN_UP,
LLM_TENSOR_DEC_OUTPUT_NORM,
// T5GEMMA specific post layer normalization tensors
LLM_TENSOR_DEC_POST_SELF_ATTN_NORM,
LLM_TENSOR_DEC_POST_CROSS_ATTN_NORM,
LLM_TENSOR_DEC_POST_FFN_NORM,
LLM_TENSOR_ENC_POST_SELF_ATTN_NORM,
LLM_TENSOR_ENC_POST_FFN_NORM,
LLM_TENSOR_ENC_ATTN_NORM,
LLM_TENSOR_ENC_ATTN_Q,
LLM_TENSOR_ENC_ATTN_K,
Expand Down
4 changes: 2 additions & 2 deletions src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -915,7 +915,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
}

// TODO: hacky solution
if (model.arch == LLM_ARCH_T5 && t_embd) {
if ((model.arch == LLM_ARCH_T5 || model.arch == LLM_ARCH_T5GEMMA) && t_embd) {
//cross.t_embd = t_embd;

synchronize();
Expand Down Expand Up @@ -1271,7 +1271,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
bool has_embd = cparams.embeddings;

// TODO: hacky enc-dec support
if (model.arch == LLM_ARCH_T5) {
if (model.arch == LLM_ARCH_T5 || model.arch == LLM_ARCH_T5GEMMA) {
has_logits = true;
has_embd = true;
}
Expand Down
Loading
Loading