model: add Hy3 (hy_v3) support with MTP speculative decoding (#25395)
* model: add Hy3 (hy_v3) architecture support Adds Tencent Hunyuan 3 (HF architecture HYV3ForCausalLM, GGUF arch hy_v3): a MoE decoder stack with per-head Q/K RMSNorm, a sigmoid router with expert selection bias, an always-active ungated shared expert, and leading dense block(s) (first_k_dense_replace). The base implementation is ported from charlie12345's fork (https://github.com/charlie12345/ROCmFPX, src/models/hyv3.cpp), adapted to current mainline APIs (hparams.n_layer(), build_qkv, build_moe_ffn with fused gate_up + scale tensors, output_s). Note: blk.N.exp_probs_b is stored without a .bias suffix for compatibility with existing hy_v3 GGUFs produced by that fork. Co-Authored-By: charlie12345 <charlie12345@users.noreply.github.com> Co-authored-by: Piotr Wilkin <ilintar@gmail.com> Assisted-by: Claude Fable 5
This commit is contained in:
co-authored by
charlie12345
Piotr Wilkin
parent
6eddde06a4
commit
2969d6d15d
@@ -750,11 +750,50 @@ const func_builtins & value_string_t::get_builtins() const {
|
||||
res->val_str.mark_input_based_on(args.get_pos(0)->val_str);
|
||||
return res;
|
||||
}},
|
||||
{"format", [](const func_args & args) -> value {
|
||||
value val_input = args.get_pos(0);
|
||||
if (!is_val<value_string>(val_input)) {
|
||||
throw raised_exception("format() first argument must be a string");
|
||||
}
|
||||
const jinja::string & fmt = val_input->as_string();
|
||||
const bool fmt_is_input = fmt.all_parts_are_input();
|
||||
|
||||
const std::string str = fmt.str();
|
||||
jinja::string result;
|
||||
std::string literal;
|
||||
auto flush_literal = [&]() {
|
||||
if (!literal.empty()) {
|
||||
result.parts.push_back({fmt_is_input, literal});
|
||||
literal.clear();
|
||||
}
|
||||
};
|
||||
|
||||
size_t arg_idx = 1; // positional args follow the format string
|
||||
for (size_t i = 0; i < str.size(); ++i) {
|
||||
if (str[i] != '{') {
|
||||
literal += str[i];
|
||||
continue;
|
||||
}
|
||||
if (i + 1 >= str.size() || str[i + 1] != '}') {
|
||||
throw not_implemented_exception("format() only supports simple '{}' placeholders");
|
||||
}
|
||||
++i;
|
||||
flush_literal();
|
||||
const jinja::string arg_str = args.get_pos(arg_idx++)->as_string();
|
||||
result.parts.insert(result.parts.end(), arg_str.parts.begin(), arg_str.parts.end());
|
||||
}
|
||||
flush_literal();
|
||||
return mk_val<value_string>(result);
|
||||
}},
|
||||
{"int", [](const func_args & args) -> value {
|
||||
value val_input = args.get_pos(0);
|
||||
value val_default = args.get_kwarg_or_pos("default", 1);
|
||||
value val_base = args.get_kwarg_or_pos("base", 2);
|
||||
const int base = val_base->is_undefined() ? 10 : val_base->as_int();
|
||||
if (base != 0 && (base < 2 || base > 36)) {
|
||||
// an out-of-range base makes std::stoi fail fast on the MSVC CRT instead of throwing
|
||||
throw raised_exception("int() base must be 0 or between 2 and 36");
|
||||
}
|
||||
if (is_val<value_string>(val_input) == false) {
|
||||
throw raised_exception("int() first argument must be a string");
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user