Skip to content

fix gemma4_assistant: arch name, tensor names, MTP dimension alignment #72

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
limitlessstoner wants to merge 2 commits into
base: feature/turboquant-kv-cache
Choose a base branch
from
Open

fix gemma4_assistant: arch name, tensor names, MTP dimension alignment #72

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
efd98f3
fix gemma4_assistant: arch name, tensor names, MTP dimension alignment
limitlessstoner Jul 30, 2026
35973a9
common: add MTP flag aliases for published-Atomic CLI compatibility
limitlessstoner Aug 12, 2026
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
10 changes: 5 additions & 5 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3877,7 +3877,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_N_GPU_LAYERS_DRAFT"));
add_opt(common_arg(
{"--spec-draft-model", "-md", "--model-draft"}, "FNAME",
{"--spec-draft-model", "-md", "--model-draft", "--mtp-head"}, "FNAME",
"draft model for speculative decoding (default: unused)",
[](common_params & params, const std::string & value) {
params.speculative.draft.mparams.path = value;
Expand Down Expand Up @@ -4023,10 +4023,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
//

add_opt(common_arg(
{"--draft", "--draft-n", "--draft-max"}, "N",
"the argument has been removed. use --spec-draft-n-max or --spec-ngram-mod-n-max",
[](common_params & /*params*/, int /*value*/) {
arg_removed("use --spec-draft-n-max or --spec-ngram-mod-n-max");
{"--draft", "--draft-n", "--draft-max", "--draft-block-size"}, "N",
"alias for --spec-draft-n-max: max number of tokens to draft per step",
[](common_params & params, int value) {
params.speculative.draft.n_max = value;
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_DRAFT_MAX"));
add_opt(common_arg(
Expand Down
1 change: 1 addition & 0 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
{"draft-simple", COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE},
{"draft-eagle3", COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3},
{"draft-mtp", COMMON_SPECULATIVE_TYPE_DRAFT_MTP},
{"mtp", COMMON_SPECULATIVE_TYPE_DRAFT_MTP},
{"draft-dflash", COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH},
{"ngram-simple", COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
{"ngram-map-k", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
Expand Down
6 changes: 3 additions & 3 deletions src/llama-arch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GEMMA3, "gemma3" },
{ LLM_ARCH_GEMMA3N, "gemma3n" },
{ LLM_ARCH_GEMMA4, "gemma4" },
{ LLM_ARCH_GEMMA4_ASSISTANT, "gemma4-assistant" },
{ LLM_ARCH_GEMMA4_ASSISTANT, "gemma4_assistant" },
{ LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
{ LLM_ARCH_STARCODER2, "starcoder2" },
{ LLM_ARCH_MAMBA, "mamba" },
Expand Down Expand Up @@ -504,8 +504,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_FFN_NORM_EXPS, "blk.%d.ffn_norm_exps" },
{ LLM_TENSOR_ATTN_K_B, "blk.%d.attn_k_b" },
{ LLM_TENSOR_ATTN_V_B, "blk.%d.attn_v_b" },
{ LLM_TENSOR_NEXTN_PROJ_PRE, "nextn.pre_projection" },
{ LLM_TENSOR_NEXTN_PROJ_POST, "nextn.post_projection" },
{ LLM_TENSOR_NEXTN_PROJ_PRE, "mtp.pre_projection" },
{ LLM_TENSOR_NEXTN_PROJ_POST, "mtp.post_projection" },
{ LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" },
{ LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
{ LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" },
Expand Down
2 changes: 1 addition & 1 deletion src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1493,7 +1493,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
}
}
}
ml.done_getting_tensors();
ml.done_getting_tensors(arch == LLM_ARCH_GEMMA4_ASSISTANT);

// Tied NVFP4 output is valid when no separate LM-head scale tensors are present.
// If sidecar scales exist, the output weight must be an actual output tensor.
Expand Down
20 changes: 14 additions & 6 deletions src/models/gemma4-assistant.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,9 @@ void llama_model_gemma4_assistant::load_arch_hparams(llama_model_loader & ml) {
hparams.f_attention_scale = 1.0f;

ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
GGML_ASSERT(hparams.n_layer_nextn == hparams.n_layer_all && "n_layer_nextn must be == n_layer_impl");
if (hparams.n_layer_nextn == 0) {
hparams.n_layer_nextn = hparams.n_layer();
}

ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
Expand All @@ -21,7 +23,7 @@ void llama_model_gemma4_assistant::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, hparams.n_embd_head_v_swa);
}

void llama_model_gemma4_assistant::load_arch_tensors(llama_model_loader &) {
void llama_model_gemma4_assistant::load_arch_tensors(llama_model_loader & ml) {
LLAMA_LOAD_LOCALS;

if (n_embd_head_k != n_embd_head_v) {
Expand All @@ -30,9 +32,6 @@ void llama_model_gemma4_assistant::load_arch_tensors(llama_model_loader &) {
if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {
throw std::runtime_error("Gemma 4 assistant requires n_embd_head_k_swa == n_embd_head_v_swa");
}
if (hparams.n_embd_out() == n_embd) {
throw std::runtime_error("Gemma 4 assistant requires embedding_length_out to carry the target hidden size");
}

tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
Expand All @@ -42,7 +41,16 @@ void llama_model_gemma4_assistant::load_arch_tensors(llama_model_loader &) {
create_tensor(tn(LLM_TENSOR_MASKED_EMBD_CENTROIDS, "weight"), {}, TENSOR_NOT_REQUIRED);
create_tensor(tn(LLM_TENSOR_MASKED_EMBD_ORDERING), {}, TENSOR_NOT_REQUIRED);

const int64_t n_embd_backbone = hparams.n_embd_inp();
// Determine backbone hidden size from projection tensor shape
int64_t n_embd_backbone = hparams.n_embd_inp();
{
auto * meta = ml.get_tensor_meta(tn(LLM_TENSOR_NEXTN_PROJ_POST, "weight").str().c_str());
if (meta && meta->ne[1] > 0) {
n_embd_backbone = meta->ne[1];
}
}
hparams.n_embd_inp_impl = n_embd_backbone;
hparams.n_embd_out_impl = n_embd_backbone;
nextn_proj_post = create_tensor(tn(LLM_TENSOR_NEXTN_PROJ_POST, "weight"), { n_embd, n_embd_backbone }, 0);

int rope_freqs_flag = 0;
Expand Down