nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #1963 from jkyamog/minimax-m3-support 

Add preliminary MiniMax-M3 support
This commit is contained in:
Kawrakow 2026-06-15 10:16:10 +02:00 committed by GitHub
parent 4f1ec69ae5 c08d194edd
commit 567854aeab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
11 changed files with 341 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

169
models/templates/MiniMax-M3.jinja Normal file
View File

@ -0,0 +1,169 @@
{# MiniMax-M3 override.
Keep MiniMax-M2's PEG-compatible tool-call wrapper, but use M3 thinking tags. #}
{%- set toolcall_begin_token = '<minimax:tool_call>' -%}
{%- set toolcall_end_token = '</minimax:tool_call>' -%}
{%- set think_begin_token = '<mm:think>' -%}
{%- set think_end_token = '</mm:think>' -%}
{#- Tool Rendering Functions ============================================== -#}
{%- macro render_tool_namespace(namespace_name, tool_list) -%}
{%- for tool in tool_list -%}
<tool>{{ tool.function | tojson(ensure_ascii=False) }}</tool>
{% endfor -%}
{%- endmacro -%}
{%- macro visible_text(content) -%}
{%- if content is string -%}
{{ content }}
{%- elif content is iterable and content is not mapping -%}
{%- for item in content -%}
{%- if item is mapping and item.type == 'text' -%}
{{- item.text }}
{%- elif item is string -%}
{{- item }}
{%- endif -%}
{%- endfor -%}
{%- else -%}
{{- content }}
{%- endif -%}
{%- endmacro -%}
{#- System Message Construction ============================================ -#}
{%- macro build_system_message(system_message) -%}
{%- if system_message and system_message.content -%}
{{- visible_text(system_message.content) }}
{%- else -%}
{%- if model_identity is not defined -%}
{%- set model_identity = "You are a helpful assistant." -%}
{%- endif -%}
{{- model_identity }}
{%- endif -%}
{%- if system_message and system_message.current_date -%}
{{- '\n' ~ 'Current date: ' + system_message.current_date }}
{%- endif -%}
{%- if system_message and system_message.current_location -%}
{{- '\n' ~ 'Current location: ' + system_message.current_location }}
{%- endif -%}
{%- endmacro -%}
{#- Main Template Logic ===================================================== -#}
{%- set system_message = none -%}
{%- set conversation_messages = messages -%}
{%- if messages and messages[0].role == "system" -%}
{%- set system_message = messages[0] -%}
{%- set conversation_messages = messages[1:] -%}
{%- endif -%}
{#- Get the last user message turn, for interleaved thinking -#}
{%- set ns = namespace(last_user_index=-1) %}
{% for m in conversation_messages %}
{%- if m.role == 'user' %}
{% set ns.last_user_index = loop.index0 -%}
{%- endif %}
{%- endfor %}
{#- Render system message -#}
{{- ']~!b[' ~ ']~b]system' ~ '\n' }}
{{- build_system_message(system_message) }}
{#- Render tools if available -#}
{%- if tools -%}
{{- '\n\n' ~ '# Tools' ~ '\n' ~ 'You may call one or more tools to assist with the user query.\nHere are the tools available in JSONSchema format:' ~ '\n' }}
{{- '\n' ~ '<tools>' ~ '\n' }}
{{- render_tool_namespace("functions", tools) }}
{{- '</tools>' ~ '\n\n' }}
{{- 'When making tool calls, use XML format to invoke tools and pass parameters:' ~ '\n' }}
{{- '\n' ~ toolcall_begin_token }}
<invoke name="tool-name">
<parameter name="param-name1">param-value-1</parameter>
<parameter name="param-name2">param-value-2</parameter>
...
</invoke>
{{- '\n' ~ toolcall_end_token }}
{%- endif -%}
{{- '[e~[\n' }}
{#- Render messages -#}
{%- set last_tool_call = namespace(name=none) -%}
{%- for message in conversation_messages -%}
{%- if message.role == 'assistant' -%}
{{- ']~b]ai' ~ '\n' }}
{%- set reasoning_content = '' %}
{%- set content = visible_text(message.content) %}
{%- if message.reasoning_content is string %}
{%- set reasoning_content = message.reasoning_content %}
{%- else %}
{%- if think_end_token in content %}
{%- set reasoning_content = content.split(think_end_token)[0].strip('\n').split(think_begin_token)[-1].strip('\n') %}
{%- set content = content.split(think_end_token)[-1].strip('\n') %}
{%- elif '</think>' in content %}
{%- set reasoning_content = content.split('</think>')[0].strip('\n').split('<think>')[-1].strip('\n') %}
{%- set content = content.split('</think>')[-1].strip('\n') %}
{%- endif %}
{%- endif %}
{%- if reasoning_content and loop.index0 > ns.last_user_index -%}
{{- think_begin_token ~ '\n' ~ reasoning_content ~ '\n' ~ think_end_token ~ '\n\n' }}
{%- endif -%}
{%- if content -%}
{{- content }}
{%- endif -%}
{%- if message.tool_calls -%}
{{- '\n' ~ toolcall_begin_token ~ '\n' }}
{%- for tool_call in message.tool_calls -%}
{%- if tool_call.function %}
{%- set tool_call = tool_call.function %}
{%- endif %}
{{- '<invoke name="' ~ tool_call.name ~ '">' }}
{% set _args = tool_call.arguments %}
{%- for k, v in _args.items() %}
{{- '<parameter name="' ~ k ~ '">' }}
{{- v | tojson(ensure_ascii=False) if v is not string else v }}
{{- '</parameter>' }}
{% endfor %}
{{- '</invoke>' ~ '\n' }}
{%- endfor -%}
{{- toolcall_end_token}}
{%- set last_tool_call.name = message.tool_calls[-1].function.name -%}
{%- else -%}
{%- set last_tool_call.name = none -%}
{%- endif -%}
{{- '[e~[' ~ '\n' }}
{%- elif message.role == 'tool' -%}
{%- if last_tool_call.name is none -%}
{{- raise_exception("Message has tool role, but there was no previous assistant message with a tool call!") }}
{%- endif -%}
{%- if loop.first or (conversation_messages[loop.index0 - 1].role != 'tool') -%}
{{- ']~b]tool' }}
{%- endif -%}
{%- if message.content is string -%}
{{- '\n' }}
{{- message.content }}
{{- '</' ~ last_tool_call.name ~ '>' }}
{%- else -%}
{%- for tr in message.content -%}
{{- '\n' }}
{{- tr.output if tr.output is defined else (tr.text if tr.type == 'text' and tr.text is defined else tr) }}
{{- '\n' }}
{%- endfor -%}
{%- endif -%}
{%- if loop.last or (conversation_messages[loop.index0 + 1].role != 'tool') -%}
{{- '[e~[\n' -}}
{%- endif -%}
{%- elif message.role == 'user' -%}
{{- ']~b]user' ~ '\n' }}
{{- visible_text(message.content) }}
{{- '[e~[' ~ '\n' }}
{%- endif -%}
{%- endfor -%}
{#- Generation prompt -#}
{%- if add_generation_prompt -%}
{{- ']~b]ai' ~ '\n' ~ think_begin_token ~ '\n' }}
{%- endif -%}

1
src/CMakeLists.txt
View File

@ -120,6 +120,7 @@ add_library(llama
graphs/build_openai.cpp
graphs/build_bailingmoe2.cpp
graphs/build_minimaxm2.cpp
graphs/build_minimaxm3.cpp
graphs/build_smollm3.cpp
)

69
src/graphs/build_minimaxm3.cpp Normal file
View File

@ -0,0 +1,69 @@
#include "../llama-build-context.h"
#include "../llama-model.h"
#include "../llama-context.h"
ggml_cgraph* llm_build_context::build_minimaxm3() {
ggml_cgraph * gf = new_graph_custom();
const int64_t n_embd_head = hparams.n_embd_head_v(0);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = n_tokens > 1 ? build_inp_out_ids() : nullptr;
ggml_tensor * KQ_mask = build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * ffn_inp = build_std_attention(gf, model.layers[il].attn_norm, inpL,
inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr, nullptr,
KQ_mask, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), 0.0f, 0,
il, true, false, true);
if ((uint32_t) il < hparams.n_layer_dense_lead) {
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
model.layers[il].ffn_up, nullptr, nullptr,
model.layers[il].ffn_gate, nullptr, nullptr,
model.layers[il].ffn_down, nullptr, nullptr,
nullptr,
LLM_FFN_SWIGLU_OAI, LLM_FFN_PAR, cb, il, gf, true);
} else {
cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
model.layers[il].ffn_gate_inp,
nullptr,
model.layers[il].ffn_up_exps,
nullptr,
model.layers[il].ffn_gate_exps,
nullptr,
model.layers[il].ffn_down_exps,
nullptr,
model.layers[il].ffn_exp_probs_b,
model.layers[il].ffn_up_shexp,
nullptr,
model.layers[il].ffn_gate_shexp,
nullptr,
model.layers[il].ffn_down_shexp,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SWIGLU_OAI_MOE,
hparams.expert_weights_norm,
hparams.expert_weights_scale != 0.0f, hparams.expert_weights_scale,
(llm_expert_gating_func_type) hparams.expert_gating_func,
LLM_FFN_SWIGLU_OAI,
cb, il, gf, true);
}
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);
inpL = cur;
}
cur = build_output(lctx, ctx0, inpL, model.output, model.output_norm, cb);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}

1
src/llama-arch.cpp
View File

@ -72,6 +72,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
{ LLM_ARCH_BAILINGMOE2, "bailingmoe2" },
{ LLM_ARCH_MINIMAX_M2, "minimax-m2" },
{ LLM_ARCH_MINIMAX_M3, "minimax-m3" },
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_MISTRAL3, "mistral3" },
{ LLM_ARCH_MIMO2, "mimo2" },

1
src/llama-arch.h
View File

@ -71,6 +71,7 @@ enum llm_arch {
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_BAILINGMOE2,
LLM_ARCH_MINIMAX_M2,
LLM_ARCH_MINIMAX_M3,
LLM_ARCH_SMOLLM3,
LLM_ARCH_MISTRAL3,
LLM_ARCH_MIMO2,

16
src/llama-build-context.cpp
View File

@ -962,6 +962,14 @@ ggml_tensor * llm_build_context::llm_build_ffn(
cur = ggml_swiglu(ctx, cur);
cb(cur, "ffn_swiglu", il);
} break;
case LLM_FFN_SWIGLU_OAI:
{
constexpr float alpha = 1.702f;
constexpr float limit = 7.0f;
cur = ggml_swiglu_oai(ctx, cur, tmp, alpha, limit);
cb(cur, "ffn_swiglu_oai", il);
type_gate = LLM_FFN_SEQ;
} break;
default:
GGML_ABORT("fatal error");
}
@ -1149,7 +1157,7 @@ llm_expert_gating_func_type gating_op,
ggml_tensor * par;
if (can_use_fmoe && up_gate_exps) {
if (up_gate_exps_b) {
if (up_gate_exps_b || type_op == LLM_FFN_SWIGLU_OAI_MOE) {
par = ggml_moe_up_gate_ext(ctx, up_gate_exps, nullptr, cur, selected_experts, up_gate_exps_b, nullptr,
type_op == LLM_FFN_SILU ? GGML_UNARY_OP_SILU :
type_op == LLM_FFN_GELU ? GGML_UNARY_OP_GELU : GGML_UNARY_OP_SWIGLU_OAI);
@ -1165,7 +1173,7 @@ llm_expert_gating_func_type gating_op,
GGML_ASSERT(!up_gate_exps && !up_gate_exps_b);
if (can_use_fmoe && lctx.cparams.fused_moe_up_gate && up_exps->type == gate_exps->type) {
if (up_exps_b || gate_exps_b) {
if (up_exps_b || gate_exps_b || type_op == LLM_FFN_SWIGLU_OAI_MOE) {
par = ggml_moe_up_gate_ext(ctx, up_exps, gate_exps, cur, selected_experts, up_exps_b, gate_exps_b,
type_op == LLM_FFN_SILU ? GGML_UNARY_OP_SILU :
type_op == LLM_FFN_GELU ? GGML_UNARY_OP_GELU : GGML_UNARY_OP_SWIGLU_OAI);
@ -2520,6 +2528,10 @@ ggml_cgraph * llm_build_context::llama_build_graph(
{
result = llm.build_minimaxm2();
} break;
case LLM_ARCH_MINIMAX_M3:
{
result = llm.build_minimaxm3();
} break;
case LLM_ARCH_SMOLLM3:
{
result = llm.build_smollm3();

2
src/llama-build-context.h
View File

@ -24,6 +24,7 @@ enum llm_ffn_op_type {
LLM_FFN_RELU,
LLM_FFN_RELU_SQR,
LLM_FFN_SWIGLU,
LLM_FFN_SWIGLU_OAI,
LLM_FFN_SWIGLU_OAI_MOE,
};
@ -312,6 +313,7 @@ struct llm_build_context {
ggml_cgraph * build_bailingmoe2();
ggml_cgraph * build_minimaxm2();
ggml_cgraph * build_minimaxm3();
ggml_cgraph * build_smollm3();

12
src/llama-hparams.cpp
View File

@ -1280,6 +1280,18 @@ void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_MINIMAX_M3:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
model.type = e_model::MODEL_UNKNOWN;
} break;
case LLM_ARCH_SMOLLM3:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);

43
src/llama-load-tensors.cpp
View File

@ -150,6 +150,7 @@ struct create_tensors_helper : public create_tensors_helper_interface {
bool create_bailingmoe2_tensors(const LLM_TN & tn);
bool create_minimaxm2_tensors(const LLM_TN & tn);
bool create_minimaxm3_tensors(const LLM_TN & tn);
bool create_smollm3_tensors(const LLM_TN & tn);
@ -3601,6 +3602,46 @@ bool create_tensors_helper::create_minimaxm2_tensors(const LLM_TN & tn) {
return use_mmap_buffer;
}
bool create_tensors_helper::create_minimaxm3_tensors(const LLM_TN & tn) {
LOADING_PRELUDE
const int64_t n_expert_shared = hparams.n_expert_shared;
const int64_t n_ff_exp = hparams.n_ff_exp;
create_embd_output(tn, n_embd, n_vocab);
for (int i = 0; i < n_layer; ++i) {
ggml_context* ctx_split = ctx_for_layer_split(i);
auto& layer = model.layers[i];
layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
layer.wk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0);
layer.wv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0);
layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
layer.attn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
layer.attn_q_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
layer.ffn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
if (i < (int) hparams.n_layer_dense_lead) {
layer.ffn_gate = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);
layer.ffn_down = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
layer.ffn_up = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
} else {
layer.ffn_gate_inp = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
layer.ffn_exp_probs_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, 0);
use_mmap_buffer &= !create_std_ffn_exps(n_embd, tn, i, 0, n_ff_exp);
layer.ffn_gate_shexp = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_exp * n_expert_shared }, 0);
layer.ffn_down_shexp = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd }, 0);
layer.ffn_up_shexp = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_exp * n_expert_shared }, 0);
}
}
return use_mmap_buffer;
}
bool create_tensors_helper::create_smollm3_tensors(const LLM_TN & tn) {
LOADING_PRELUDE
@ -4412,6 +4453,8 @@ bool create_tensors_helper::create_tensors() {
use_mmap_buffer = create_bailingmoe2_tensors(tn); break;
case LLM_ARCH_MINIMAX_M2:
use_mmap_buffer = create_minimaxm2_tensors(tn); break;
case LLM_ARCH_MINIMAX_M3:
use_mmap_buffer = create_minimaxm3_tensors(tn); break;
case LLM_ARCH_SMOLLM3:
use_mmap_buffer = create_smollm3_tensors(tn); break;
case LLM_ARCH_MIMO2:

27
src/llama-model.cpp
View File

@ -1532,6 +1532,33 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
},
},
{
LLM_ARCH_MINIMAX_M3,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
{ LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
},
},
{
LLM_ARCH_SMOLLM3,
{

2
src/llama.cpp
View File

@ -3071,6 +3071,7 @@ static bool is_model_split_supported(const llama_model & model) {
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_ERNIE4_5_MOE,
LLM_ARCH_MINIMAX_M2,
LLM_ARCH_MINIMAX_M3,
LLM_ARCH_SEED_OSS,
LLM_ARCH_STEP35,
LLM_ARCH_LAGUNA,
@ -7439,6 +7440,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_OPENAI_MOE:
case LLM_ARCH_BAILINGMOE2:
case LLM_ARCH_MINIMAX_M2:
case LLM_ARCH_MINIMAX_M3:
case LLM_ARCH_MIMO2:
case LLM_ARCH_SEED_OSS:
case LLM_ARCH_STEP35: