mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-05-19 08:00:25 +00:00
Merge commit '3e4bb29666' into concedo_experimental
# Conflicts: # .github/workflows/build.yml # ci/run.sh # cmake/common.cmake # examples/eval-callback/CMakeLists.txt # examples/model-conversion/scripts/causal/modelcard.template # ggml/src/ggml-cuda/fattn.cu # ggml/src/ggml-metal/CMakeLists.txt # src/CMakeLists.txt # tests/CMakeLists.txt # tests/test-arg-parser.cpp
This commit is contained in:
Concedo
commit
af7811dbe1
26 changed files with 1072 additions and 166 deletions
|
|
@ -1403,6 +1403,118 @@ static void common_chat_parse_solar_open(common_chat_msg_parser & builder) {
|
|||
builder.add_content(builder.consume_rest());
|
||||
}
|
||||
|
||||
static void common_chat_parse_exaone_moe_content(common_chat_msg_parser & builder) {
|
||||
// 1) <tool_call>{ "name": "...", "arguments": {...} }</tool_call>
|
||||
// 2) <tool_call>{ "id": "...", "type": "function", "function": { "name": "...", "arguments": {...} } }</tool_call>
|
||||
static const common_regex tool_call_open(R"(<tool_call[^>]*>)");
|
||||
|
||||
if (!builder.syntax().parse_tool_calls) {
|
||||
LOG_DBG("%s: not parse_tool_calls\n", __func__);
|
||||
builder.add_content(builder.consume_rest());
|
||||
return;
|
||||
}
|
||||
|
||||
LOG_DBG("%s: parse_tool_calls\n", __func__);
|
||||
|
||||
// Find all <tool_call></tool_call> blocks
|
||||
while (auto first = builder.try_find_regex(tool_call_open, std::string::npos, /* add_prelude_to_content= */ true)) {
|
||||
builder.move_to(first->groups[0].end);
|
||||
builder.consume_spaces();
|
||||
|
||||
builder.try_consume_literal("```json");
|
||||
builder.try_consume_literal("```");
|
||||
builder.consume_spaces();
|
||||
|
||||
// Consume JSON object
|
||||
auto data = builder.consume_json();
|
||||
|
||||
builder.consume_spaces();
|
||||
builder.try_consume_literal("```");
|
||||
builder.consume_spaces();
|
||||
|
||||
if (!builder.try_consume_literal("</tool_call>")) {
|
||||
throw common_chat_msg_partial_exception("incomplete tool call");
|
||||
}
|
||||
builder.consume_spaces();
|
||||
|
||||
// Extract name and arguments
|
||||
std::string name;
|
||||
std::string id;
|
||||
nlohmann::ordered_json arguments;
|
||||
|
||||
const auto extract_args = [&](const nlohmann::ordered_json & obj) -> bool {
|
||||
if (!obj.contains("name") || !obj.contains("arguments")) {
|
||||
return false;
|
||||
}
|
||||
name = obj.at("name").get<std::string>();
|
||||
arguments = obj.at("arguments");
|
||||
if (obj.contains("id") && obj.at("id").is_string()) {
|
||||
id = obj.at("id").get<std::string>();
|
||||
}
|
||||
return true;
|
||||
};
|
||||
|
||||
if (!extract_args(data.json)) {
|
||||
if (data.json.contains("function") && data.json.at("function").is_object()) {
|
||||
auto fn = data.json.at("function");
|
||||
extract_args(fn);
|
||||
if (id.empty() && data.json.contains("id") && data.json.at("id").is_string()) {
|
||||
id = data.json.at("id").get<std::string>();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// If name is empty, treat the JSON object as content
|
||||
if (name.empty()) {
|
||||
LOG_DBG("%s: tool call missing name, treating as content\n", __func__);
|
||||
builder.add_content(data.json.dump());
|
||||
continue;
|
||||
}
|
||||
|
||||
std::string args_str = arguments.dump();
|
||||
if (!builder.add_tool_call(name, id, args_str)) {
|
||||
throw common_chat_msg_partial_exception("incomplete tool call");
|
||||
}
|
||||
}
|
||||
|
||||
builder.add_content(builder.consume_rest());
|
||||
}
|
||||
|
||||
static void common_chat_parse_exaone_moe(common_chat_msg_parser & builder) {
|
||||
LOG_DBG("%s: parsing exaone_moe\n", __func__);
|
||||
// EXAONE MoE outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
|
||||
// First try to parse using the standard reasoning parsing method
|
||||
LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
|
||||
|
||||
auto start_pos = builder.pos();
|
||||
auto found_end_think = builder.try_find_literal("</think>");
|
||||
builder.move_to(start_pos);
|
||||
|
||||
if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
|
||||
LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
|
||||
common_chat_parse_exaone_moe_content(builder);
|
||||
} else if (builder.try_parse_reasoning("<think>", "</think>")) {
|
||||
// If reasoning was parsed successfully, the remaining content is regular content
|
||||
LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
|
||||
common_chat_parse_exaone_moe_content(builder);
|
||||
} else {
|
||||
if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
|
||||
LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
|
||||
common_chat_parse_exaone_moe_content(builder);
|
||||
return;
|
||||
}
|
||||
// If no reasoning tags found, check if we should treat everything as reasoning
|
||||
if (builder.syntax().thinking_forced_open) {
|
||||
// If thinking is forced open but no tags found, treat everything as reasoning
|
||||
LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
|
||||
builder.add_reasoning_content(builder.consume_rest());
|
||||
} else {
|
||||
LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
|
||||
common_chat_parse_exaone_moe_content(builder);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
|
||||
builder.try_parse_reasoning("<think>", "</think>");
|
||||
builder.add_content(builder.consume_rest());
|
||||
|
|
@ -1490,6 +1602,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
|
|||
case COMMON_CHAT_FORMAT_SOLAR_OPEN:
|
||||
common_chat_parse_solar_open(builder);
|
||||
break;
|
||||
case COMMON_CHAT_FORMAT_EXAONE_MOE:
|
||||
common_chat_parse_exaone_moe(builder);
|
||||
break;
|
||||
default:
|
||||
throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
|
||||
}
|
||||
|
|
|
|||
|
|
@ -676,6 +676,7 @@ const char * common_chat_format_name(common_chat_format format) {
|
|||
case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
|
||||
case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
|
||||
case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
|
||||
case COMMON_CHAT_FORMAT_EXAONE_MOE: return "EXAONE MoE";
|
||||
case COMMON_CHAT_FORMAT_PEG_SIMPLE: return "peg-simple";
|
||||
case COMMON_CHAT_FORMAT_PEG_NATIVE: return "peg-native";
|
||||
case COMMON_CHAT_FORMAT_PEG_CONSTRUCTED: return "peg-constructed";
|
||||
|
|
@ -2545,6 +2546,65 @@ static common_chat_params common_chat_params_init_solar_open(const common_chat_t
|
|||
return data;
|
||||
}
|
||||
|
||||
static common_chat_params common_chat_params_init_exaone_moe(const common_chat_template & tmpl, const struct templates_params & inputs) {
|
||||
common_chat_params data;
|
||||
|
||||
data.prompt = apply(tmpl, inputs);
|
||||
data.format = COMMON_CHAT_FORMAT_EXAONE_MOE;
|
||||
if (string_ends_with(data.prompt, "<think>\n")) {
|
||||
if (!inputs.enable_thinking) {
|
||||
data.prompt += "</think>\n\n";
|
||||
} else {
|
||||
data.thinking_forced_open = true;
|
||||
}
|
||||
}
|
||||
|
||||
if (inputs.tools.is_array() && !inputs.tools.empty()) {
|
||||
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null();
|
||||
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
|
||||
std::vector<std::string> tool_rules;
|
||||
foreach_function(inputs.tools, [&](const json & tool) {
|
||||
const auto & function = tool.at("function");
|
||||
std::string name = function.at("name");
|
||||
auto parameters = function.at("parameters");
|
||||
builder.resolve_refs(parameters);
|
||||
// Expect: <tool_call>{"name": "<name>", "arguments": {...}}</tool_call>
|
||||
tool_rules.push_back(builder.add_rule(
|
||||
name + "-call",
|
||||
"\"<tool_call>\" space " +
|
||||
builder.add_schema(name + "-obj", json{
|
||||
{"type", "object"},
|
||||
{"properties", {
|
||||
{"name", json{{"const", name}}},
|
||||
{"arguments", parameters},
|
||||
}},
|
||||
{"required", json::array({"name", "arguments"})},
|
||||
}) +
|
||||
" space \"</tool_call>\" space"));
|
||||
});
|
||||
|
||||
auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | "));
|
||||
builder.add_rule("root",
|
||||
std::string(data.thinking_forced_open ? "( \"</think>\" space )? " : "") +
|
||||
(inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call));
|
||||
|
||||
data.grammar_triggers.push_back({
|
||||
COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
|
||||
std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)?" : "") +
|
||||
"(<tool_call>)[\\s\\S]*"
|
||||
});
|
||||
data.preserved_tokens = {
|
||||
"<think>",
|
||||
"</think>",
|
||||
"<tool_call>",
|
||||
"</tool_call>",
|
||||
};
|
||||
});
|
||||
}
|
||||
|
||||
return data;
|
||||
}
|
||||
|
||||
static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
|
||||
common_chat_params data;
|
||||
data.prompt = apply(tmpl, inputs);
|
||||
|
|
@ -2715,6 +2775,13 @@ static common_chat_params common_chat_templates_apply_jinja(
|
|||
return common_chat_params_init_xiaomi_mimo(tmpl, params);
|
||||
}
|
||||
|
||||
// EXAONE MoE format detection
|
||||
if (src.find("<tool_call>") != std::string::npos &&
|
||||
src.find("<tool_result>") != std::string::npos &&
|
||||
src.find("<|tool_declare|>") != std::string::npos) {
|
||||
return common_chat_params_init_exaone_moe(tmpl, params);
|
||||
}
|
||||
|
||||
// Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
|
||||
if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
|
||||
return common_chat_params_init_hermes_2_pro(tmpl, params);
|
||||
|
|
|
|||
|
|
@ -125,6 +125,7 @@ enum common_chat_format {
|
|||
COMMON_CHAT_FORMAT_APRIEL_1_5,
|
||||
COMMON_CHAT_FORMAT_XIAOMI_MIMO,
|
||||
COMMON_CHAT_FORMAT_SOLAR_OPEN,
|
||||
COMMON_CHAT_FORMAT_EXAONE_MOE,
|
||||
|
||||
// These are intended to be parsed by the PEG parser
|
||||
COMMON_CHAT_FORMAT_PEG_SIMPLE,
|
||||
|
|
|
|||
|
|
@ -1252,6 +1252,9 @@ class TextModel(ModelBase):
|
|||
if chkhsh == "16389f0a1f51ee53e562ffd51c371dc508639ab0e4261502071836e50e223e91":
|
||||
# ref: https://huggingface.co/upstage/Solar-Open-100B
|
||||
res = "solar-open"
|
||||
if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
|
||||
# ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
|
||||
res = "exaone-moe"
|
||||
|
||||
if res is None:
|
||||
logger.warning("\n")
|
||||
|
|
@ -8748,6 +8751,106 @@ class Exaone4Model(TextModel):
|
|||
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
|
||||
|
||||
|
||||
@ModelBase.register("ExaoneMoEForCausalLM")
|
||||
class ExaoneMoEModel(Exaone4Model):
|
||||
model_arch = gguf.MODEL_ARCH.EXAONE_MOE
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
|
||||
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
super().set_gguf_parameters()
|
||||
self.gguf_writer.add_expert_count(self.hparams["num_experts"])
|
||||
moe_intermediate_size = self.hparams["moe_intermediate_size"]
|
||||
num_shared_experts = self.hparams["num_shared_experts"]
|
||||
self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
|
||||
self.gguf_writer.add_expert_shared_count(num_shared_experts)
|
||||
self.gguf_writer.add_expert_shared_feed_forward_length(moe_intermediate_size * num_shared_experts)
|
||||
self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
|
||||
self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
|
||||
n_dense_layer = self.hparams.get("first_k_dense_replace", self.hparams.get("first_last_k_dense_replace", 0))
|
||||
self.gguf_writer.add_leading_dense_block_count(n_dense_layer)
|
||||
# For here, we hard-code the number of NextN/MTP layers to 1 for K-EXAONE,
|
||||
# so that we can convert MTP weights to GGUF format for speculative decoding.
|
||||
# This is because HF config of K-EXAONE does not have `num_nextn_predict_layers` at now.
|
||||
# Will be updated when HF config is updated.
|
||||
self.gguf_writer.add_nextn_predict_layers(self.hparams.get("num_nextn_predict_layers", 1))
|
||||
|
||||
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
|
||||
|
||||
_experts: list[dict[str, Tensor]] | None = None
|
||||
|
||||
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
||||
if name.startswith("mtp."):
|
||||
if name.find("layers.") != -1:
|
||||
# `mtp.layers.0.[module_name]` format
|
||||
name = name.replace(f"mtp.layers.{bid}", f"model.layers.{bid + self.hparams['num_hidden_layers']}")
|
||||
else:
|
||||
# mtp fc/norm weights
|
||||
remapper = {
|
||||
"mtp.fc": "model.layers.{bid}.eh_proj",
|
||||
"mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
|
||||
"mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm",
|
||||
"mtp.norm": "model.layers.{bid}.shared_head.norm",
|
||||
}
|
||||
_n = Path(name)
|
||||
new_name = remapper[_n.stem] + _n.suffix
|
||||
|
||||
# set shared weights for all NextN/MTP layers
|
||||
tensors = []
|
||||
for bid in range(self.hparams['num_hidden_layers'], self.block_count):
|
||||
new_name = new_name.format(bid=bid)
|
||||
tensors.append((self.map_tensor_name(new_name), data_torch))
|
||||
return tensors
|
||||
|
||||
if name.endswith("e_score_correction_bias"):
|
||||
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
|
||||
|
||||
if name.find("mlp.experts") != -1:
|
||||
n_experts = self.hparams["num_experts"]
|
||||
assert bid is not None
|
||||
|
||||
if self._experts is None:
|
||||
self._experts = [{} for _ in range(self.block_count)]
|
||||
|
||||
self._experts[bid][name] = data_torch
|
||||
|
||||
if len(self._experts[bid]) >= n_experts * 3:
|
||||
tensors: list[tuple[str, Tensor]] = []
|
||||
|
||||
# merge the experts into a single 3d tensor
|
||||
for w_name in ["down_proj", "gate_proj", "up_proj"]:
|
||||
datas: list[Tensor] = []
|
||||
|
||||
for xid in range(n_experts):
|
||||
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
|
||||
datas.append(self._experts[bid][ename])
|
||||
del self._experts[bid][ename]
|
||||
|
||||
data_torch = torch.stack(datas, dim=0)
|
||||
|
||||
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
|
||||
|
||||
new_name = self.map_tensor_name(merged_name)
|
||||
|
||||
tensors.append((new_name, data_torch))
|
||||
return tensors
|
||||
else:
|
||||
return []
|
||||
|
||||
return [(self.map_tensor_name(name), data_torch)]
|
||||
|
||||
def prepare_tensors(self):
|
||||
super().prepare_tensors()
|
||||
if self._experts is not None:
|
||||
# flatten `list[dict[str, Tensor]]` into `list[str]`
|
||||
experts = [k for d in self._experts for k in d.keys()]
|
||||
if len(experts) > 0:
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
|
||||
@ModelBase.register("GraniteForCausalLM")
|
||||
class GraniteModel(LlamaModel):
|
||||
"""Conversion for IBM's GraniteForCausalLM"""
|
||||
|
|
|
|||
|
|
@ -147,6 +147,7 @@ models = [
|
|||
{"name": "kormo", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/KORMo-Team/KORMo-tokenizer", },
|
||||
{"name": "youtu", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
|
||||
{"name": "solar-open", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
|
||||
{"name": "exaone-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
|
||||
]
|
||||
|
||||
# some models are known to be broken upstream, so we will skip them as exceptions
|
||||
|
|
|
|||
|
|
@ -267,6 +267,10 @@ static const char * cu_get_error_str(CUresult err) {
|
|||
#define FLASH_ATTN_AVAILABLE
|
||||
#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
|
||||
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
#define LDMATRIX_TRANS_AVAILABLE
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
|
||||
static bool fp16_available(const int cc) {
|
||||
return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
|
||||
(GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
|
||||
|
|
|
|||
|
|
@ -914,7 +914,7 @@ void launch_fattn(
|
|||
|
||||
const int nblocks_stream_k = max_blocks;
|
||||
|
||||
const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || tiles_efficiency_percent < 75;
|
||||
const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;
|
||||
|
||||
blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
|
||||
blocks_num.y = 1;
|
||||
|
|
|
|||
|
|
@ -98,6 +98,19 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
|
|||
return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
|
||||
}
|
||||
|
||||
static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_rdna(const int DKQ, const int DV, const int ncols) {
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 16, 128, 2, 64, 128, 128, 128, 2, true);
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2, 64, 128, 128, 64, 2, true);
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2, 64, 128, 128, 64, 2, true);
|
||||
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16, 64, 4, 32, 96, 64, 128, 1, false);
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2, 32, 160, 128, 128, 1, false);
|
||||
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1, 32, 160, 128, 128, 1, false);
|
||||
|
||||
// TODO tune specifically for RDNA
|
||||
return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
|
||||
}
|
||||
|
||||
static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
|
||||
if (ampere_mma_available(cc)) {
|
||||
return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
|
||||
|
|
@ -105,6 +118,9 @@ static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, c
|
|||
if (turing_mma_available(cc)) {
|
||||
return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
|
||||
}
|
||||
if (amd_wmma_available(cc)) {
|
||||
return ggml_cuda_fattn_mma_get_config_rdna(DKQ, DV, ncols);
|
||||
}
|
||||
GGML_ASSERT(volta_mma_available(cc));
|
||||
return ggml_cuda_fattn_mma_get_config_volta(DKQ, DV, ncols);
|
||||
}
|
||||
|
|
@ -116,6 +132,8 @@ static constexpr __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config(cons
|
|||
return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
|
||||
#elif defined(VOLTA_MMA_AVAILABLE)
|
||||
return ggml_cuda_fattn_mma_get_config_volta(DKQ, DV, ncols);
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
return ggml_cuda_fattn_mma_get_config_rdna(DKQ, DV, ncols);
|
||||
#else
|
||||
GGML_UNUSED_VARS(DKQ, DV, ncols);
|
||||
return fattn_mma_config(32, 1, 0, 0, 0, 0, 0, false);
|
||||
|
|
@ -186,6 +204,23 @@ static constexpr __device__ bool ggml_cuda_fattn_mma_get_Q_in_reg(const int DKQ,
|
|||
return ggml_cuda_fattn_mma_get_config(DKQ, DV, ncols).Q_in_reg;
|
||||
}
|
||||
|
||||
static constexpr __device__ int get_cols_per_thread() {
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
return 1; // RDNA has a single column.
|
||||
#else
|
||||
return 2; // This is specifically KQ columns, Volta only has a single VKQ column.
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
}
|
||||
|
||||
static __host__ int get_cols_per_warp(const int cc) {
|
||||
if (turing_mma_available(cc) || amd_wmma_available(cc)) {
|
||||
return 16;
|
||||
} else {
|
||||
// Volta
|
||||
return 32;
|
||||
}
|
||||
}
|
||||
|
||||
// ------------------------------------------------------------------------------------------------------------------
|
||||
|
||||
static __host__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, const int DV, const int ncols1, const int ncols2, const int cc) {
|
||||
|
|
@ -393,10 +428,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
const int jt,
|
||||
const int kb0,
|
||||
const int k_VKQ_sup) {
|
||||
#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
|
||||
#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
|
||||
constexpr int ncols = ncols1 * ncols2;
|
||||
constexpr int cols_per_warp = T_B_KQ::I;
|
||||
constexpr int cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
|
||||
constexpr int cols_per_thread = get_cols_per_thread();
|
||||
constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
|
||||
constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
|
||||
constexpr int nbatch_K2 = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
|
||||
|
|
@ -413,6 +448,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
const int k_VKQ_0 = kb0 * nbatch_fa;
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
T_C_KQ KQ_C[nbatch_fa/(np*(cols_per_warp == 8 ? T_C_KQ::I : T_C_KQ::J))];
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
|
||||
#else // Volta
|
||||
T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
|
|
@ -461,8 +498,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
if constexpr (cols_per_warp == 8) {
|
||||
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
|
||||
} else {
|
||||
// Wide version of KQ_C is column-major => swap A and B.
|
||||
// Wide version of KQ_C is column-major
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
// RDNA matrix C is column-major.
|
||||
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
|
||||
#else
|
||||
// swap A and B for CUDA.
|
||||
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[k_KQ_0/T_A_KQ::J], K_A);
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -479,8 +522,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
T_A_KQ K_A;
|
||||
load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
|
||||
|
||||
// Wide version of KQ_C is column-major => swap A and B.
|
||||
// Wide version of KQ_C is column-major
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
// RDNA matrix C is column-major.
|
||||
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
|
||||
#else
|
||||
// swap A and B for CUDA.
|
||||
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -532,7 +581,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
#pragma unroll
|
||||
for (int l = 0; l < T_C_KQ::ne; ++l) {
|
||||
if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
|
||||
KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k0/(np*T_C_KQ::I)].x[l] + FATTN_KQ_MAX_OFFSET);
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
constexpr int KQ_idx = 0;
|
||||
#else
|
||||
// Turing + Volta:
|
||||
const int KQ_idx = l % 2;
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[k0/(np*T_C_KQ::I)].x[l] + FATTN_KQ_MAX_OFFSET);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -552,8 +607,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
#pragma unroll
|
||||
for (int l = 0; l < T_C_KQ::ne; ++l) {
|
||||
if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
|
||||
KQ_C[k0/(np*T_C_KQ::I)].x[l] = expf(KQ_C[k0/(np*T_C_KQ::I)].x[l] - KQ_max_new[l % 2]);
|
||||
KQ_rowsum_add[l % 2] += KQ_C[k0/(np*T_C_KQ::I)].x[l];
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
constexpr int KQ_idx = 0;
|
||||
#else
|
||||
// Turing + Volta:
|
||||
const int KQ_idx = l % 2;
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
KQ_C[k0/(np*T_C_KQ::I)].x[l] = expf(KQ_C[k0/(np*T_C_KQ::I)].x[l] - KQ_max_new[KQ_idx]);
|
||||
KQ_rowsum_add[KQ_idx] += KQ_C[k0/(np*T_C_KQ::I)].x[l];
|
||||
} else {
|
||||
KQ_C[k0/(np*T_C_KQ::I)].x[l] = 0.0f;
|
||||
}
|
||||
|
|
@ -584,8 +645,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
#pragma unroll
|
||||
for (int l = 0; l < T_C_KQ::ne; ++l) {
|
||||
if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
constexpr int KQ_idx = 0;
|
||||
#else
|
||||
// Turing + Volta:
|
||||
KQ_max_new[(l/2) % 2] = fmaxf(KQ_max_new[(l/2) % 2], KQ_C[(k0/(np*T_C_KQ::J))].x[l] + FATTN_KQ_MAX_OFFSET);
|
||||
const int KQ_idx = (l/2) % 2;
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[(k0/(np*T_C_KQ::J))].x[l] + FATTN_KQ_MAX_OFFSET);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -596,7 +662,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
// Values per KQ column are spread across 4 threads:
|
||||
constexpr int offset_first = 2;
|
||||
constexpr int offset_last = 1;
|
||||
#else
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
// Values per KQ column are spread across 2 threads:
|
||||
constexpr int offset_first = 16;
|
||||
constexpr int offset_last = 16;
|
||||
#else // Volta
|
||||
// Values per KQ column are spread across 2 threads:
|
||||
constexpr int offset_first = 2;
|
||||
constexpr int offset_last = 2;
|
||||
|
|
@ -612,10 +682,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
for (int k0 = 0; k0 < nbatch_fa; k0 += np*T_C_KQ::J) {
|
||||
#pragma unroll
|
||||
for (int l = 0; l < T_C_KQ::ne; ++l) {
|
||||
// Turing + Volta:
|
||||
if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
|
||||
KQ_C[(k0/(np*T_C_KQ::J))].x[l] = expf(KQ_C[(k0/(np*T_C_KQ::J))].x[l] - KQ_max_new[(l/2) % 2]);
|
||||
KQ_rowsum_add[(l/2) % 2] += KQ_C[(k0/(np*T_C_KQ::J))].x[l];
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
constexpr int KQ_idx = 0;
|
||||
#else
|
||||
// Turing + Volta:
|
||||
const int KQ_idx = (l/2) % 2;
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
KQ_C[(k0/(np*T_C_KQ::J))].x[l] = expf(KQ_C[(k0/(np*T_C_KQ::J))].x[l] - KQ_max_new[KQ_idx]);
|
||||
KQ_rowsum_add[KQ_idx] += KQ_C[(k0/(np*T_C_KQ::J))].x[l];
|
||||
} else {
|
||||
KQ_C[(k0/(np*T_C_KQ::J))].x[l] = 0.0f;
|
||||
}
|
||||
|
|
@ -639,7 +714,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
if constexpr (cols_per_warp == 8) {
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
|
||||
#pragma unroll
|
||||
for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
|
||||
#pragma unroll
|
||||
|
|
@ -660,6 +735,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
}
|
||||
}
|
||||
}
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
const half2 KQ_max_scale_h2 = make_half2(
|
||||
KQ_max_scale[0], KQ_max_scale[0]);
|
||||
#pragma unroll
|
||||
for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
|
||||
#pragma unroll
|
||||
for (int l = 0; l < T_C_VKQ::ne; ++l) {
|
||||
VKQ_C[i].x[l] *= KQ_max_scale_h2;
|
||||
}
|
||||
}
|
||||
#else // Volta
|
||||
const half2 KQ_max_scale_h2 = make_half2(
|
||||
KQ_max_scale[(threadIdx.x / 2) % 2], KQ_max_scale[(threadIdx.x / 2) % 2]);
|
||||
|
|
@ -707,6 +792,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
// Therefore, iterate over V in reverse and re-use the data if possible.
|
||||
static_assert(!mla || nstages <= 1, "combination of MLA and multi-stage loading not implemented");
|
||||
constexpr int reusable_cutoff = mla ? (DKQ - 1) - (DKQ - 1) % (2*nbatch_K2) - (DKQ - DV) : DV;
|
||||
#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
|
||||
T_A_VKQ A_identity;
|
||||
make_identity_mat(A_identity);
|
||||
#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
|
||||
|
||||
// Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
|
||||
#pragma unroll
|
||||
|
|
@ -727,7 +816,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
}
|
||||
const half2 * tile_V_i = i0_start < reusable_cutoff ? tile_V : tile_V + (i0_start - reusable_cutoff)/2;
|
||||
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
|
||||
constexpr int i0_stride = cols_per_warp == 8 ? T_C_VKQ::I : 2*T_C_VKQ::J;
|
||||
#pragma unroll
|
||||
for (int i_VKQ_0 = i0_start; i_VKQ_0 < i0_stop; i_VKQ_0 += i0_stride) {
|
||||
|
|
@ -737,12 +826,26 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
|
||||
|
||||
T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
|
||||
#if defined(LDMATRIX_TRANS_AVAILABLE)
|
||||
load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
|
||||
#else
|
||||
// TODO: Try to transpose tile_V when loading gmem to smem.
|
||||
// Use mma to transpose T_A_VKQ for RDNA.
|
||||
T_A_VKQ A_trans;
|
||||
load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
|
||||
mma(A, A_trans, A_identity);
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
if constexpr (T_B_KQ::I == 8) {
|
||||
mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
|
||||
} else {
|
||||
// Wide version of VKQ_C is column-major => swap A and B.
|
||||
// Wide version of VKQ_C is column-major.
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
// RDNA matrix C is column-major.
|
||||
mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
|
||||
#else
|
||||
// swap A and B for CUDA.
|
||||
mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::J)], A);
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -761,7 +864,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::I)], A);
|
||||
}
|
||||
}
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
#endif // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
|
||||
|
||||
if constexpr (nstages <= 1) {
|
||||
__syncthreads(); // Only needed if tile_K == tile_V.
|
||||
|
|
@ -774,7 +877,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
|
|||
tile_Q, tile_K, tile_V, tile_mask,
|
||||
Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
|
||||
NO_DEVICE_CODE;
|
||||
#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
|
||||
#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
|
||||
}
|
||||
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
|
|
@ -794,6 +897,15 @@ template<> struct mma_tile_sizes<8> {
|
|||
using T_B_VKQ = tile< 8, 8, half2>; // column-major
|
||||
using T_C_VKQ = tile<16, 4, half2>; // row-major
|
||||
};
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
template<int ncols> struct mma_tile_sizes {
|
||||
using T_A_KQ = tile<16, 8, half2>; // row-major
|
||||
using T_B_KQ = tile<16, 8, half2>; // column-major
|
||||
using T_C_KQ = tile<16, 16, float>; // column-major
|
||||
using T_A_VKQ = tile<16, 8, half2>; // row-major
|
||||
using T_B_VKQ = tile<16, 8, half2>; // column-major
|
||||
using T_C_VKQ = tile<16, 8, half2>; // column-major
|
||||
};
|
||||
#else // Volta
|
||||
template<int ncols> struct mma_tile_sizes {
|
||||
using T_A_KQ = tile< 8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
|
||||
|
|
@ -828,7 +940,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
const int jt,
|
||||
const int kb0_start,
|
||||
const int kb0_stop) {
|
||||
#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
|
||||
#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
|
||||
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
|
||||
|
||||
constexpr int ncols = ncols1 * ncols2;
|
||||
|
|
@ -840,7 +952,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
using T_C_VKQ = typename mma_tile_sizes<ncols>::T_C_VKQ;
|
||||
|
||||
constexpr int cols_per_warp = T_B_KQ::I;
|
||||
constexpr int cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
|
||||
constexpr int cols_per_thread = get_cols_per_thread();
|
||||
constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
|
||||
constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa (DKQ, DV, ncols);
|
||||
constexpr int nbatch_K2 = ggml_cuda_fattn_mma_get_nbatch_K2 (DKQ, DV, ncols);
|
||||
|
|
@ -871,6 +983,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
T_B_KQ Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
T_C_VKQ VKQ_C[ DV/(2*T_C_VKQ::J)];
|
||||
#else // Volta
|
||||
T_C_VKQ VKQ_C[ DV/(2*T_C_VKQ::J)];
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
|
|
@ -1010,6 +1124,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
// The partial sums are spread across 8/4 threads.
|
||||
constexpr int offset_first = cols_per_warp == 8 ? 16 : 2;
|
||||
constexpr int offset_last = cols_per_warp == 8 ? 4 : 1;
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
// The partial sums are spread across 2 threads.
|
||||
constexpr int offset_first = 16;
|
||||
constexpr int offset_last = 16;
|
||||
#else // Volta
|
||||
// The partial sums are spread across 2 threads.
|
||||
constexpr int offset_first = 2;
|
||||
|
|
@ -1047,7 +1165,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
|
||||
#if defined(TURING_MMA_AVAILABLE)
|
||||
if constexpr (cols_per_warp == 8) {
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
|
||||
#pragma unroll
|
||||
for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
|
||||
#pragma unroll
|
||||
|
|
@ -1068,6 +1186,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
}
|
||||
}
|
||||
}
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[0]);
|
||||
#pragma unroll
|
||||
for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
|
||||
#pragma unroll
|
||||
for (int l = 0; l < T_C_VKQ::ne; ++l) {
|
||||
VKQ_C[i].x[l] *= KQ_max_scale_h2;
|
||||
}
|
||||
}
|
||||
#else // Volta
|
||||
const int col = (threadIdx.x / 2) % 2;
|
||||
const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
|
||||
|
|
@ -1119,6 +1246,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(threadIdx.x % 4);
|
||||
const float2 KQ_cmr = make_float2(KQ_max[threadIdx.x % cols_per_thread], KQ_rowsum[threadIdx.x % cols_per_thread]);
|
||||
const bool thread_should_write = threadIdx.x % 4 < cols_per_thread;
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(0);
|
||||
const float2 KQ_cmr = make_float2(KQ_max[0], KQ_rowsum[0]);
|
||||
const bool thread_should_write = threadIdx.x / 16 < cols_per_thread;
|
||||
#else // Volta
|
||||
const int jc_cwm = threadIdx.y*cols_per_warp + T_C_KQ::get_i(threadIdx.x & 2);
|
||||
const float2 KQ_cmr = make_float2(KQ_max[(threadIdx.x & 2) / 2], KQ_rowsum[(threadIdx.x & 2) / 2]);
|
||||
|
|
@ -1319,7 +1450,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
|
||||
jt, kb0_start, kb0_stop);
|
||||
NO_DEVICE_CODE;
|
||||
#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
|
||||
#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
|
||||
}
|
||||
|
||||
template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool mla>
|
||||
|
|
@ -1346,7 +1477,7 @@ static __global__ void flash_attn_ext_f16(
|
|||
const int32_t nb21, const int32_t nb22, const int64_t nb23,
|
||||
const int32_t ne31, const int32_t ne32, const int32_t ne33,
|
||||
const int32_t nb31, const int32_t nb32, const int64_t nb33) {
|
||||
#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
|
||||
#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
|
||||
|
||||
// Skip unused kernel variants for faster compilation:
|
||||
if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
|
||||
|
|
@ -1360,6 +1491,13 @@ static __global__ void flash_attn_ext_f16(
|
|||
}
|
||||
#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
|
||||
|
||||
#if defined(AMD_WMMA_AVAILABLE)
|
||||
if (ncols1*ncols2 > 32 || ncols1*ncols2 < 16 || DKQ > 128 || ncols2 == 1) {
|
||||
NO_DEVICE_CODE;
|
||||
return;
|
||||
}
|
||||
#endif // defined(AMD_WMMA_AVAILABLE)
|
||||
|
||||
static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
|
||||
|
||||
constexpr int ncols = ncols1 * ncols2;
|
||||
|
|
@ -1473,7 +1611,7 @@ static __global__ void flash_attn_ext_f16(
|
|||
ne31, ne32, ne33,
|
||||
nb31, nb32, nb33);
|
||||
NO_DEVICE_CODE;
|
||||
#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
|
||||
#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
|
||||
}
|
||||
|
||||
template <int DKQ, int DV, int ncols1, int ncols2>
|
||||
|
|
@ -1492,7 +1630,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
|
|||
const bool Q_in_reg = ggml_cuda_fattn_mma_get_Q_in_reg (DKQ, DV, ncols, cc);
|
||||
const int nstages = ggml_cuda_fattn_mma_get_nstages (DKQ, DV, ncols1, ncols2, cc);
|
||||
|
||||
const int cols_per_warp = std::min(ncols, turing_mma_available(cc) ? 16 : 32);
|
||||
const int cols_per_warp = std::min(ncols, get_cols_per_warp(cc));
|
||||
const int nwarps = nthreads / WARP_SIZE;
|
||||
|
||||
constexpr bool mla = DKQ == 576;
|
||||
|
|
@ -1512,29 +1650,34 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
|
|||
float logit_softcap;
|
||||
memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
|
||||
|
||||
#if defined(GGML_USE_HIP)
|
||||
using fattn_kernel_ptr_t = const void*;
|
||||
#else
|
||||
using fattn_kernel_ptr_t = fattn_kernel_t;
|
||||
#endif // defined(GGML_USE_HIP)
|
||||
fattn_kernel_t fattn_kernel;
|
||||
if (logit_softcap == 0.0f) {
|
||||
constexpr bool use_logit_softcap = false;
|
||||
fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
|
||||
|
||||
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
||||
#if !defined(GGML_USE_MUSA)
|
||||
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
|
||||
if (!shared_memory_limit_raised[id]) {
|
||||
CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
|
||||
CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
|
||||
shared_memory_limit_raised[id] = true;
|
||||
}
|
||||
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
||||
#endif // !defined(GGML_USE_MUSA)
|
||||
} else {
|
||||
constexpr bool use_logit_softcap = true;
|
||||
fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
|
||||
|
||||
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
||||
#if !defined(GGML_USE_MUSA)
|
||||
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
|
||||
if (!shared_memory_limit_raised[id]) {
|
||||
CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
|
||||
CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
|
||||
shared_memory_limit_raised[id] = true;
|
||||
}
|
||||
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
|
||||
#endif // !defined(GGML_USE_MUSA)
|
||||
}
|
||||
|
||||
launch_fattn<DV, ncols1, ncols2>
|
||||
|
|
|
|||
|
|
@ -10,7 +10,7 @@ static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
|
|||
return 128;
|
||||
}
|
||||
|
||||
// Currenlty llvm with the amdgcn target dose not support unrolling loops
|
||||
// Currenlty llvm with the amdgcn target does not support unrolling loops
|
||||
// that contain a break that can not be resolved at compile time.
|
||||
#ifdef __clang__
|
||||
#pragma clang diagnostic push
|
||||
|
|
|
|||
|
|
@ -18,12 +18,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
|
|||
}
|
||||
}
|
||||
|
||||
if (turing_mma_available(cc) && Q->ne[1] <= 16/ncols2) {
|
||||
if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
|
||||
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
|
||||
return;
|
||||
}
|
||||
|
||||
if (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING || Q->ne[1] <= 32/ncols2) {
|
||||
if (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING || amd_wmma_available(cc) || Q->ne[1] <= 32/ncols2) {
|
||||
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
|
||||
return;
|
||||
}
|
||||
|
|
@ -230,7 +230,18 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
|
|||
|
||||
// The effective batch size for the kernel can be increased by gqa_ratio.
|
||||
// The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
|
||||
const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
|
||||
bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
|
||||
for (const ggml_tensor * t : {Q, K, V, mask}) {
|
||||
if (t == nullptr) {
|
||||
continue;
|
||||
}
|
||||
for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
|
||||
if (t->nb[i] % 16 != 0) {
|
||||
gqa_opt_applies = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const int cc = ggml_cuda_info().devices[device].cc;
|
||||
|
||||
|
|
@ -324,6 +335,31 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
|
|||
}
|
||||
#endif // defined(GGML_HIP_ROCWMMA_FATTN)
|
||||
|
||||
if (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc) && gqa_opt_applies && Q->ne[0] <= 128 && Q->ne[0] != 40 && Q->ne[0] != 72) {
|
||||
if (can_use_vector_kernel) {
|
||||
if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
|
||||
if (Q->ne[1] == 1) {
|
||||
if (!gqa_opt_applies) {
|
||||
return BEST_FATTN_KERNEL_VEC;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if (Q->ne[1] <= 2) {
|
||||
return BEST_FATTN_KERNEL_VEC;
|
||||
}
|
||||
}
|
||||
}
|
||||
int gqa_ratio_eff = 1;
|
||||
const int ncols2_max = Q->ne[0] == 576 ? 16 : 8;
|
||||
while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
|
||||
gqa_ratio_eff *= 2;
|
||||
}
|
||||
if (Q->ne[1] * gqa_ratio_eff <= 8) {
|
||||
return BEST_FATTN_KERNEL_TILE; // AMD WMMA is only faster if the full tile width of 16 can be utilized.
|
||||
}
|
||||
return BEST_FATTN_KERNEL_MMA_F16;
|
||||
}
|
||||
|
||||
// If there are no tensor cores available, use the generic tile kernel:
|
||||
if (can_use_vector_kernel) {
|
||||
if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
|
||||
|
|
|
|||
|
|
@ -206,10 +206,16 @@ namespace ggml_cuda_mma {
|
|||
|
||||
static __device__ __forceinline__ int get_j(const int l) {
|
||||
if constexpr (I == 16 && J == 16) {
|
||||
// matrix C
|
||||
#if defined(RDNA3)
|
||||
return 2 * l + (threadIdx.x / 16);
|
||||
if constexpr (std::is_same_v<T, float> || std::is_same_v<T, int>) {
|
||||
// matrix C
|
||||
return 2 * l + (threadIdx.x / 16);
|
||||
} else {
|
||||
// matrix A&B
|
||||
return l;
|
||||
}
|
||||
#else
|
||||
// matrix C is the transposed matrix A&B on RDNA4
|
||||
return ne * (threadIdx.x / 16) + l;
|
||||
#endif // defined(RDNA3)
|
||||
} else if constexpr (I == 16 && J == 8) {
|
||||
|
|
@ -621,6 +627,21 @@ namespace ggml_cuda_mma {
|
|||
|
||||
return ret;
|
||||
}
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
template <int I, int J>
|
||||
static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
|
||||
tile<I, J/2, half2> ret;
|
||||
#pragma unroll
|
||||
for (int l0 = 0; l0 < tile_float.ne; l0 += 2) {
|
||||
ret.x[l0/2] = make_half2(tile_float.x[l0 + 0], tile_float.x[l0 + 1]);
|
||||
}
|
||||
return ret;
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ tile<8, 8, half2> get_transposed(const tile<16, 4, half2> & t) {
|
||||
NO_DEVICE_CODE;
|
||||
return tile<8, 8, half2>{};
|
||||
}
|
||||
#else // Volta
|
||||
template <int I, int J>
|
||||
static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
|
||||
|
|
@ -639,6 +660,19 @@ namespace ggml_cuda_mma {
|
|||
}
|
||||
#endif // defined(TURING_MMA_AVAILABLE)
|
||||
|
||||
static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
|
||||
#if defined(RDNA4)
|
||||
const int row = t.get_i(0);
|
||||
const int left_right = t.get_j(0) / 4;
|
||||
const int up_down = row / 8;
|
||||
const int idx = row % 8;
|
||||
reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
|
||||
#else
|
||||
GGML_UNUSED_VARS(t);
|
||||
NO_DEVICE_CODE;
|
||||
#endif // defined(RDNA4)
|
||||
}
|
||||
|
||||
template <int I, int J, typename T, data_layout dl>
|
||||
static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
|
||||
#if defined(AMD_MFMA_AVAILABLE)
|
||||
|
|
@ -878,6 +912,17 @@ namespace ggml_cuda_mma {
|
|||
: "+r"(Dxi[2]), "+r"(Dxi[3])
|
||||
: "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
|
||||
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
|
||||
#elif defined(AMD_WMMA_AVAILABLE)
|
||||
#if defined(RDNA4)
|
||||
using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
|
||||
halfx8_t& acc_frag = reinterpret_cast<halfx8_t&>(D.x[0]);
|
||||
const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
|
||||
const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
|
||||
acc_frag = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
|
||||
#else
|
||||
GGML_UNUSED_VARS(D, A, B);
|
||||
NO_DEVICE_CODE;
|
||||
#endif // defined(RDNA4)
|
||||
#else
|
||||
GGML_UNUSED_VARS(D, A, B);
|
||||
NO_DEVICE_CODE;
|
||||
|
|
|
|||
2
ggml/src/ggml-cuda/vendors/hip.h
vendored
2
ggml/src/ggml-cuda/vendors/hip.h
vendored
|
|
@ -138,6 +138,8 @@
|
|||
#define cudaStream_t hipStream_t
|
||||
#define cudaSuccess hipSuccess
|
||||
#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
|
||||
#define cudaFuncSetAttribute hipFuncSetAttribute
|
||||
#define cudaFuncAttributeMaxDynamicSharedMemorySize hipFuncAttributeMaxDynamicSharedMemorySize
|
||||
#define __trap() do { abort(); __builtin_unreachable(); } while(0)
|
||||
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
|
||||
#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
|
||||
|
|
|
|||
|
|
@ -14451,13 +14451,29 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
|
|||
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
|
||||
const vk_device& device = ggml_vk_get_device(ctx->device);
|
||||
|
||||
const bool uses_bda = (op->op == GGML_OP_IM2COL || op->op == GGML_OP_IM2COL_3D) &&
|
||||
device->shader_int64 && device->buffer_device_address;
|
||||
|
||||
auto const & tensor_size_supported = [&](size_t tensor_size) {
|
||||
if (tensor_size > device->max_buffer_size) {
|
||||
return false;
|
||||
}
|
||||
// For im2col shaders using BDA, maxStorageBufferRange limit doesn't apply.
|
||||
// If shader64BitIndexing is enabled, maxStorageBufferRange limit doesn't apply.
|
||||
if (!uses_bda && !device->shader_64b_indexing) {
|
||||
if (tensor_size > device->properties.limits.maxStorageBufferRange) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
return true;
|
||||
};
|
||||
// reject any tensors larger than the max buffer size
|
||||
for (int i = 0; i < GGML_MAX_SRC; i++) {
|
||||
if (op->src[i] && ggml_nbytes(op->src[i]) > device->max_buffer_size) {
|
||||
if (op->src[i] && !tensor_size_supported(ggml_nbytes(op->src[i]))) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
if (ggml_nbytes(op) > device->max_buffer_size) {
|
||||
if (!tensor_size_supported(ggml_nbytes(op))) {
|
||||
return false;
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -264,7 +264,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
|
|||
const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8 ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
|
||||
(((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
|
||||
|
||||
buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
|
||||
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -334,7 +334,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
|
|||
(data_a[ib_k].scales[is+4] >> 4) | ((data_a[ib_k].scales[is ] & 0xC0) >> 2));
|
||||
}
|
||||
|
||||
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
|
||||
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -385,7 +385,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
|
|||
const uint is = iqs_k / 4;
|
||||
const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
|
||||
|
||||
buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
|
||||
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -424,6 +424,7 @@ class MODEL_ARCH(IntEnum):
|
|||
NEMOTRON_H_MOE = auto()
|
||||
EXAONE = auto()
|
||||
EXAONE4 = auto()
|
||||
EXAONE_MOE = auto()
|
||||
GRANITE = auto()
|
||||
GRANITE_MOE = auto()
|
||||
GRANITE_HYBRID = auto()
|
||||
|
|
@ -843,6 +844,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
|
|||
MODEL_ARCH.NEMOTRON_H_MOE: "nemotron_h_moe",
|
||||
MODEL_ARCH.EXAONE: "exaone",
|
||||
MODEL_ARCH.EXAONE4: "exaone4",
|
||||
MODEL_ARCH.EXAONE_MOE: "exaone-moe",
|
||||
MODEL_ARCH.GRANITE: "granite",
|
||||
MODEL_ARCH.GRANITE_MOE: "granitemoe",
|
||||
MODEL_ARCH.GRANITE_HYBRID: "granitehybrid",
|
||||
|
|
@ -2754,6 +2756,38 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
|||
MODEL_TENSOR.FFN_UP,
|
||||
MODEL_TENSOR.FFN_POST_NORM,
|
||||
],
|
||||
MODEL_ARCH.EXAONE_MOE: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
MODEL_TENSOR.OUTPUT,
|
||||
MODEL_TENSOR.ROPE_FREQS,
|
||||
MODEL_TENSOR.ATTN_NORM,
|
||||
MODEL_TENSOR.ATTN_Q,
|
||||
MODEL_TENSOR.ATTN_Q_NORM,
|
||||
MODEL_TENSOR.ATTN_K,
|
||||
MODEL_TENSOR.ATTN_K_NORM,
|
||||
MODEL_TENSOR.ATTN_V,
|
||||
MODEL_TENSOR.ATTN_OUT,
|
||||
MODEL_TENSOR.FFN_NORM,
|
||||
MODEL_TENSOR.FFN_GATE,
|
||||
MODEL_TENSOR.FFN_DOWN,
|
||||
MODEL_TENSOR.FFN_UP,
|
||||
MODEL_TENSOR.FFN_GATE_INP,
|
||||
MODEL_TENSOR.FFN_GATE_EXP,
|
||||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
MODEL_TENSOR.FFN_GATE_SHEXP,
|
||||
MODEL_TENSOR.FFN_DOWN_SHEXP,
|
||||
MODEL_TENSOR.FFN_UP_SHEXP,
|
||||
MODEL_TENSOR.FFN_EXP_PROBS_B,
|
||||
# NextN/MTP tensors - preserved but unused
|
||||
MODEL_TENSOR.NEXTN_EH_PROJ,
|
||||
MODEL_TENSOR.NEXTN_EMBED_TOKENS,
|
||||
MODEL_TENSOR.NEXTN_ENORM,
|
||||
MODEL_TENSOR.NEXTN_HNORM,
|
||||
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
|
||||
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
|
||||
],
|
||||
MODEL_ARCH.GRANITE: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
|
|
|
|||
|
|
@ -436,7 +436,8 @@ class TensorNameMap:
|
|||
"model.layers.{bid}.mlp.expert_bias", # afmoe
|
||||
"model.layers.{bid}.feed_forward.expert_bias", # lfm2moe
|
||||
"model.layers.{bid}.block_sparse_moe.e_score_correction", # minimax-m2
|
||||
"backbone.layers.{bid}.mixer.gate.e_score_correction" # nemotron-h-moe
|
||||
"backbone.layers.{bid}.mixer.gate.e_score_correction", # nemotron-h-moe
|
||||
"model.layers.{bid}.mlp.e_score_correction", # exaone-moe
|
||||
),
|
||||
|
||||
# Feed-forward up
|
||||
|
|
@ -1794,7 +1795,7 @@ class TensorNameMap:
|
|||
"model.embed_audio.soft_embedding_norm", # gemma3n
|
||||
),
|
||||
|
||||
# NextN/MTP tensors for GLM4_MOE
|
||||
# NextN/MTP tensors
|
||||
MODEL_TENSOR.NEXTN_EH_PROJ: (
|
||||
"model.layers.{bid}.eh_proj",
|
||||
),
|
||||
|
|
|
|||
|
|
@ -81,6 +81,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_NEMOTRON_H_MOE, "nemotron_h_moe" },
|
||||
{ LLM_ARCH_EXAONE, "exaone" },
|
||||
{ LLM_ARCH_EXAONE4, "exaone4" },
|
||||
{ LLM_ARCH_EXAONE_MOE, "exaone-moe" },
|
||||
{ LLM_ARCH_RWKV6, "rwkv6" },
|
||||
{ LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
|
||||
{ LLM_ARCH_RWKV7, "rwkv7" },
|
||||
|
|
@ -1728,6 +1729,38 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
|
|||
LLM_TENSOR_FFN_UP,
|
||||
LLM_TENSOR_FFN_POST_NORM,
|
||||
};
|
||||
case LLM_ARCH_EXAONE_MOE:
|
||||
return {
|
||||
LLM_TENSOR_TOKEN_EMBD,
|
||||
LLM_TENSOR_OUTPUT_NORM,
|
||||
LLM_TENSOR_OUTPUT,
|
||||
LLM_TENSOR_ROPE_FREQS,
|
||||
LLM_TENSOR_ATTN_NORM,
|
||||
LLM_TENSOR_ATTN_Q,
|
||||
LLM_TENSOR_ATTN_Q_NORM,
|
||||
LLM_TENSOR_ATTN_K,
|
||||
LLM_TENSOR_ATTN_K_NORM,
|
||||
LLM_TENSOR_ATTN_V,
|
||||
LLM_TENSOR_ATTN_OUT,
|
||||
LLM_TENSOR_FFN_NORM,
|
||||
LLM_TENSOR_FFN_GATE,
|
||||
LLM_TENSOR_FFN_DOWN,
|
||||
LLM_TENSOR_FFN_UP,
|
||||
LLM_TENSOR_FFN_GATE_INP,
|
||||
LLM_TENSOR_FFN_GATE_EXPS,
|
||||
LLM_TENSOR_FFN_DOWN_EXPS,
|
||||
LLM_TENSOR_FFN_UP_EXPS,
|
||||
LLM_TENSOR_FFN_GATE_SHEXP,
|
||||
LLM_TENSOR_FFN_UP_SHEXP,
|
||||
LLM_TENSOR_FFN_DOWN_SHEXP,
|
||||
LLM_TENSOR_FFN_EXP_PROBS_B,
|
||||
LLM_TENSOR_NEXTN_EH_PROJ,
|
||||
LLM_TENSOR_NEXTN_EMBED_TOKENS,
|
||||
LLM_TENSOR_NEXTN_ENORM,
|
||||
LLM_TENSOR_NEXTN_HNORM,
|
||||
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
|
||||
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
|
||||
};
|
||||
case LLM_ARCH_RWKV6:
|
||||
return {
|
||||
LLM_TENSOR_TOKEN_EMBD,
|
||||
|
|
|
|||
|
|
@ -85,6 +85,7 @@ enum llm_arch {
|
|||
LLM_ARCH_NEMOTRON_H_MOE,
|
||||
LLM_ARCH_EXAONE,
|
||||
LLM_ARCH_EXAONE4,
|
||||
LLM_ARCH_EXAONE_MOE,
|
||||
LLM_ARCH_RWKV6,
|
||||
LLM_ARCH_RWKV6QWEN2,
|
||||
LLM_ARCH_RWKV7,
|
||||
|
|
|
|||
|
|
@ -57,6 +57,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
|
|||
{ "minicpm", LLM_CHAT_TEMPLATE_MINICPM },
|
||||
{ "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 },
|
||||
{ "exaone4", LLM_CHAT_TEMPLATE_EXAONE_4 },
|
||||
{ "exaone-moe", LLM_CHAT_TEMPLATE_EXAONE_MOE },
|
||||
{ "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD },
|
||||
{ "granite", LLM_CHAT_TEMPLATE_GRANITE },
|
||||
{ "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT },
|
||||
|
|
@ -137,6 +138,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
|
|||
} else if (tmpl_contains("[gMASK]<sop>")) {
|
||||
return LLM_CHAT_TEMPLATE_CHATGLM_4;
|
||||
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
|
||||
if (tmpl_contains("<|tool_declare|>")) {
|
||||
return LLM_CHAT_TEMPLATE_EXAONE_MOE;
|
||||
}
|
||||
return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
|
||||
} else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
|
||||
return LLM_CHAT_TEMPLATE_GLMEDGE;
|
||||
|
|
@ -576,6 +580,22 @@ int32_t llm_chat_apply_template(
|
|||
if (add_ass) {
|
||||
ss << "[|assistant|]";
|
||||
}
|
||||
} else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_MOE) {
|
||||
for (auto message : chat) {
|
||||
std::string role(message->role);
|
||||
if (role == "system") {
|
||||
ss << "<|system|>\n" << trim(message->content) << "<|endofturn|>\n";
|
||||
} else if (role == "user") {
|
||||
ss << "<|user|>\n" << trim(message->content) << "<|endofturn|>\n";
|
||||
} else if (role == "assistant") {
|
||||
ss << "<|assistant|>\n" << trim(message->content) << "<|endofturn|>\n";
|
||||
} else if (role == "tool") {
|
||||
ss << "<|tool|>\n" << trim(message->content) << "<|endofturn|>\n";
|
||||
}
|
||||
}
|
||||
if (add_ass) {
|
||||
ss << "<|assistant|>\n";
|
||||
}
|
||||
} else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
|
||||
// this template requires the model to have "\n\n" as EOT token
|
||||
for (size_t i = 0; i < chat.size(); i++) {
|
||||
|
|
|
|||
|
|
@ -36,6 +36,7 @@ enum llm_chat_template {
|
|||
LLM_CHAT_TEMPLATE_MINICPM,
|
||||
LLM_CHAT_TEMPLATE_EXAONE_3,
|
||||
LLM_CHAT_TEMPLATE_EXAONE_4,
|
||||
LLM_CHAT_TEMPLATE_EXAONE_MOE,
|
||||
LLM_CHAT_TEMPLATE_RWKV_WORLD,
|
||||
LLM_CHAT_TEMPLATE_GRANITE,
|
||||
LLM_CHAT_TEMPLATE_GIGACHAT,
|
||||
|
|
|
|||
|
|
@ -244,11 +244,14 @@ struct llama_file::impl {
|
|||
}
|
||||
errno = 0;
|
||||
if (fd == -1) {
|
||||
std::size_t ret = std::fread(ptr, len, 1, fp);
|
||||
const size_t curr_off = tell();
|
||||
const size_t to_read = std::min(len, size - curr_off);
|
||||
|
||||
std::size_t ret = std::fread(ptr, to_read, 1, fp);
|
||||
if (ferror(fp)) {
|
||||
throw std::runtime_error(format("read error: %s", strerror(errno)));
|
||||
}
|
||||
if (ret != 1) {
|
||||
if (to_read > 0 && ret != 1) {
|
||||
throw std::runtime_error("unexpectedly reached end of file");
|
||||
}
|
||||
} else {
|
||||
|
|
@ -617,9 +620,9 @@ struct llama_mlock::impl {
|
|||
|
||||
char* errmsg = std::strerror(errno);
|
||||
bool suggest = (errno == ENOMEM);
|
||||
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
|
||||
// visionOS/tvOS dont't support RLIMIT_MEMLOCK
|
||||
// Skip resource limit checks on visionOS/tvOS
|
||||
#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX) || defined(__HAIKU__)
|
||||
// visionOS/tvOS/Haiku don't support RLIMIT_MEMLOCK
|
||||
// Skip resource limit checks on these platforms
|
||||
suggest = false;
|
||||
#else
|
||||
struct rlimit lock_limit;
|
||||
|
|
|
|||
|
|
@ -53,6 +53,7 @@
|
|||
#include "models/ernie4-5.cpp"
|
||||
#include "models/exaone.cpp"
|
||||
#include "models/exaone4.cpp"
|
||||
#include "models/exaone-moe.cpp"
|
||||
#include "models/falcon-h1.cpp"
|
||||
#include "models/falcon.cpp"
|
||||
#include "models/gemma-embedding.cpp"
|
||||
|
|
@ -555,7 +556,7 @@ struct llama_model::impl {
|
|||
llama_mlocks mlock_bufs;
|
||||
llama_mlocks mlock_mmaps;
|
||||
|
||||
// contexts where the model tensors metadata is stored as well ass the corresponding buffers:
|
||||
// contexts where the model tensors metadata is stored as well as the corresponding buffers:
|
||||
std::vector<std::pair<ggml_context_ptr, std::vector<ggml_backend_buffer_ptr>>> ctxs_bufs;
|
||||
|
||||
buft_list_t cpu_buft_list;
|
||||
|
|
@ -2042,6 +2043,38 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
|||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_EXAONE_MOE:
|
||||
{
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
||||
hparams.n_swa = 128;
|
||||
hparams.set_swa_pattern(4);
|
||||
hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
|
||||
hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
|
||||
|
||||
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, true);
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert);
|
||||
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used);
|
||||
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared, false);
|
||||
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
|
||||
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
|
||||
ml.get_key(LLM_KV_EXPERT_GROUP_COUNT, hparams.n_expert_groups, false);
|
||||
ml.get_key(LLM_KV_EXPERT_GROUP_USED_COUNT, hparams.n_group_used, false);
|
||||
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
|
||||
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_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
|
||||
|
||||
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
case 32: type = LLM_TYPE_30B_A3B; break;
|
||||
case 48:
|
||||
case 49: type = LLM_TYPE_235B_A22B; break;
|
||||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_RWKV6:
|
||||
case LLM_ARCH_RWKV6QWEN2:
|
||||
{
|
||||
|
|
@ -5678,6 +5711,84 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
|
|||
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_EXAONE_MOE:
|
||||
{
|
||||
const int64_t n_ff_exp = hparams.n_ff_exp;
|
||||
const int64_t n_expert = hparams.n_expert;
|
||||
const int64_t n_expert_used = hparams.n_expert_used;
|
||||
const int64_t n_ff_shexp = hparams.n_ff_shexp;
|
||||
const int64_t head_dim = hparams.n_embd_head_k;
|
||||
const int64_t n_qo_dim = n_head * head_dim;
|
||||
const int64_t n_kv_dim = n_head_kv * head_dim;
|
||||
|
||||
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
|
||||
|
||||
// output
|
||||
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
|
||||
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
|
||||
|
||||
if (output == NULL) {
|
||||
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
|
||||
}
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
int flags = 0;
|
||||
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
|
||||
// skip all tensors in the NextN layers
|
||||
flags |= TENSOR_SKIP;
|
||||
}
|
||||
|
||||
auto & layer = layers[i];
|
||||
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_qo_dim}, flags);
|
||||
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_kv_dim}, flags);
|
||||
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_kv_dim}, flags);
|
||||
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, flags);
|
||||
|
||||
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0) | flags);
|
||||
|
||||
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
|
||||
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
|
||||
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
|
||||
|
||||
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
|
||||
|
||||
// dense layers for first n_layer_dense_lead layers or nextn_predict_layers layers at the end
|
||||
if (i < (int) hparams.n_layer_dense_lead || (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers)) {
|
||||
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
|
||||
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, flags);
|
||||
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags);
|
||||
} else {
|
||||
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
|
||||
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
|
||||
|
||||
if (n_expert == 0) {
|
||||
throw std::runtime_error("n_expert must be > 0");
|
||||
}
|
||||
if (n_expert_used == 0) {
|
||||
throw std::runtime_error("n_expert_used must be > 0");
|
||||
}
|
||||
|
||||
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
|
||||
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, flags);
|
||||
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
|
||||
|
||||
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
|
||||
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
|
||||
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
|
||||
}
|
||||
|
||||
// NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
|
||||
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
|
||||
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
|
||||
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, flags);
|
||||
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, flags);
|
||||
|
||||
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
|
||||
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
|
||||
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
|
||||
}
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_RWKV6:
|
||||
{
|
||||
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
|
||||
|
|
@ -7264,59 +7375,59 @@ void llama_model::print_info() const {
|
|||
};
|
||||
|
||||
// hparams
|
||||
LLAMA_LOG_INFO("%s: arch = %s\n", __func__, arch_name().c_str());
|
||||
LLAMA_LOG_INFO("%s: vocab_only = %d\n", __func__, hparams.vocab_only);
|
||||
LLAMA_LOG_INFO("%s: no_alloc = %d\n", __func__, hparams.no_alloc);
|
||||
LLAMA_LOG_INFO("%s: arch = %s\n", __func__, arch_name().c_str());
|
||||
LLAMA_LOG_INFO("%s: vocab_only = %d\n", __func__, hparams.vocab_only);
|
||||
LLAMA_LOG_INFO("%s: no_alloc = %d\n", __func__, hparams.no_alloc);
|
||||
|
||||
if (!hparams.vocab_only) {
|
||||
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
|
||||
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
|
||||
LLAMA_LOG_INFO("%s: n_embd_inp = %u\n", __func__, hparams.n_embd_inp());
|
||||
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
|
||||
LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot);
|
||||
LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa);
|
||||
LLAMA_LOG_INFO("%s: is_swa_any = %u\n", __func__, hparams.is_swa_any());
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v);
|
||||
LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_embd_k_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_embd_v_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps);
|
||||
LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps);
|
||||
LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv);
|
||||
LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias);
|
||||
LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale);
|
||||
LLAMA_LOG_INFO("%s: f_attn_scale = %.1e\n", __func__, hparams.f_attention_scale);
|
||||
LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert);
|
||||
LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used);
|
||||
LLAMA_LOG_INFO("%s: n_expert_groups = %d\n", __func__, hparams.n_expert_groups);
|
||||
LLAMA_LOG_INFO("%s: n_group_used = %d\n", __func__, hparams.n_group_used);
|
||||
LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn);
|
||||
LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type);
|
||||
LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type);
|
||||
LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str());
|
||||
LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train);
|
||||
LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train);
|
||||
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
|
||||
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
|
||||
LLAMA_LOG_INFO("%s: n_embd_inp = %u\n", __func__, hparams.n_embd_inp());
|
||||
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
|
||||
LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot);
|
||||
LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa);
|
||||
LLAMA_LOG_INFO("%s: is_swa_any = %u\n", __func__, hparams.is_swa_any());
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v);
|
||||
LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_embd_k_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_embd_v_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps);
|
||||
LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps);
|
||||
LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv);
|
||||
LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias);
|
||||
LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale);
|
||||
LLAMA_LOG_INFO("%s: f_attn_scale = %.1e\n", __func__, hparams.f_attention_scale);
|
||||
LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert);
|
||||
LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used);
|
||||
LLAMA_LOG_INFO("%s: n_expert_groups = %d\n", __func__, hparams.n_expert_groups);
|
||||
LLAMA_LOG_INFO("%s: n_group_used = %d\n", __func__, hparams.n_group_used);
|
||||
LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn);
|
||||
LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type);
|
||||
LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type);
|
||||
LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str());
|
||||
LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train);
|
||||
LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train);
|
||||
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
|
||||
LLAMA_LOG_INFO("%s: freq_base_swa = %.1f\n", __func__, hparams.rope_freq_base_train_swa);
|
||||
LLAMA_LOG_INFO("%s: freq_scale_swa = %g\n", __func__, hparams.rope_freq_scale_train_swa);
|
||||
LLAMA_LOG_INFO("%s: freq_base_swa = %.1f\n", __func__, hparams.rope_freq_base_train_swa);
|
||||
LLAMA_LOG_INFO("%s: freq_scale_swa = %g\n", __func__, hparams.rope_freq_scale_train_swa);
|
||||
}
|
||||
LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn);
|
||||
LLAMA_LOG_INFO("%s: rope_yarn_log_mul= %.4f\n", __func__, hparams.rope_yarn_log_mul);
|
||||
LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown");
|
||||
LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn);
|
||||
LLAMA_LOG_INFO("%s: rope_yarn_log_mul = %.4f\n", __func__, hparams.rope_yarn_log_mul);
|
||||
LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown");
|
||||
// MRoPE (Multi-axis Rotary Position Embedding) sections
|
||||
if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
|
||||
LLAMA_LOG_INFO("%s: mrope sections = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]);
|
||||
LLAMA_LOG_INFO("%s: mrope sections = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]);
|
||||
}
|
||||
if (!classifier_labels.empty()) {
|
||||
LLAMA_LOG_INFO("%s: n_cls_out = %u\n", __func__, hparams.n_cls_out);
|
||||
LLAMA_LOG_INFO("%s: n_cls_out = %u\n", __func__, hparams.n_cls_out);
|
||||
|
||||
size_t i = 0;
|
||||
for (auto label : classifier_labels) {
|
||||
LLAMA_LOG_INFO("%s: cls_label[%2zu] = %s\n", __func__, i++, label.c_str());
|
||||
LLAMA_LOG_INFO("%s: cls_label[%2zu] = %s\n", __func__, i++, label.c_str());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -7330,55 +7441,55 @@ void llama_model::print_info() const {
|
|||
arch == LLM_ARCH_QWEN3NEXT ||
|
||||
arch == LLM_ARCH_NEMOTRON_H ||
|
||||
arch == LLM_ARCH_NEMOTRON_H_MOE) {
|
||||
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
|
||||
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
|
||||
LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state);
|
||||
LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank);
|
||||
LLAMA_LOG_INFO("%s: ssm_n_group = %u\n", __func__, hparams.ssm_n_group);
|
||||
LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms = %d\n", __func__, hparams.ssm_dt_b_c_rms);
|
||||
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
|
||||
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
|
||||
LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state);
|
||||
LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank);
|
||||
LLAMA_LOG_INFO("%s: ssm_n_group = %u\n", __func__, hparams.ssm_n_group);
|
||||
LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms = %d\n", __func__, hparams.ssm_dt_b_c_rms);
|
||||
}
|
||||
|
||||
LLAMA_LOG_INFO("%s: model type = %s\n", __func__, type_name().c_str());
|
||||
LLAMA_LOG_INFO("%s: model type = %s\n", __func__, type_name().c_str());
|
||||
if (pimpl->n_elements >= 1e12) {
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f T\n", __func__, pimpl->n_elements*1e-12);
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f T\n", __func__, pimpl->n_elements*1e-12);
|
||||
} else if (pimpl->n_elements >= 1e9) {
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, pimpl->n_elements*1e-9);
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, pimpl->n_elements*1e-9);
|
||||
} else if (pimpl->n_elements >= 1e6) {
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f M\n", __func__, pimpl->n_elements*1e-6);
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f M\n", __func__, pimpl->n_elements*1e-6);
|
||||
} else {
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f K\n", __func__, pimpl->n_elements*1e-3);
|
||||
LLAMA_LOG_INFO("%s: model params = %.2f K\n", __func__, pimpl->n_elements*1e-3);
|
||||
}
|
||||
|
||||
// general kv
|
||||
LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, name.c_str());
|
||||
LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, name.c_str());
|
||||
|
||||
if (arch == LLM_ARCH_DEEPSEEK) {
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_DEEPSEEK2) {
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
|
||||
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
|
||||
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_QWEN2MOE) {
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE || arch == LLM_ARCH_RND1) {
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_MINICPM ||
|
||||
|
|
@ -7386,41 +7497,41 @@ void llama_model::print_info() const {
|
|||
arch == LLM_ARCH_GRANITE_MOE ||
|
||||
arch == LLM_ARCH_GRANITE_HYBRID ||
|
||||
arch == LLM_ARCH_NEMOTRON_H_MOE) {
|
||||
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
|
||||
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
|
||||
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
|
||||
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
|
||||
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_BAILINGMOE) {
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_BAILINGMOE2) {
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
LLAMA_LOG_INFO("%s: nextn_predict_layers = %d\n", __func__, hparams.nextn_predict_layers);
|
||||
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
|
||||
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
|
||||
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
LLAMA_LOG_INFO("%s: nextn_predict_layers = %d\n", __func__, hparams.nextn_predict_layers);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_GROVEMOE) {
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_chexp = %d\n", __func__, hparams.n_ff_chexp);
|
||||
LLAMA_LOG_INFO("%s: n_group_experts = %d\n", __func__, hparams.n_group_experts);
|
||||
LLAMA_LOG_INFO("%s: expert_group_scale = %.2f\n", __func__, hparams.expert_group_scale);
|
||||
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
|
||||
LLAMA_LOG_INFO("%s: n_ff_chexp = %d\n", __func__, hparams.n_ff_chexp);
|
||||
LLAMA_LOG_INFO("%s: n_group_experts = %d\n", __func__, hparams.n_group_experts);
|
||||
LLAMA_LOG_INFO("%s: expert_group_scale = %.2f\n", __func__, hparams.expert_group_scale);
|
||||
}
|
||||
|
||||
vocab.print_info();
|
||||
|
|
@ -7977,6 +8088,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
|
|||
llm = std::make_unique<llm_build_exaone4<false>>(*this, params);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_EXAONE_MOE:
|
||||
{
|
||||
llm = std::make_unique<llm_build_exaone_moe>(*this, params);
|
||||
} break;
|
||||
case LLM_ARCH_RWKV6:
|
||||
{
|
||||
llm = std::make_unique<llm_build_rwkv6>(*this, params);
|
||||
|
|
@ -8337,6 +8452,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
|
|||
case LLM_ARCH_NEMOTRON:
|
||||
case LLM_ARCH_EXAONE:
|
||||
case LLM_ARCH_EXAONE4:
|
||||
case LLM_ARCH_EXAONE_MOE:
|
||||
case LLM_ARCH_MINICPM3:
|
||||
case LLM_ARCH_BAILINGMOE2:
|
||||
case LLM_ARCH_DOTS1:
|
||||
|
|
|
|||
|
|
@ -686,6 +686,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
|
|||
"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
|
||||
};
|
||||
break;
|
||||
case LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE:
|
||||
regex_exprs = {
|
||||
// original regex from tokenizer.json
|
||||
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?(?:\\p{L}\\p{M}*(?: \\p{L}\\p{M}*)*)+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]?|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+"
|
||||
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?(?:\\p{L}\\p{M}*(?: \\p{L}\\p{M}*)*)+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]?|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
|
||||
};
|
||||
break;
|
||||
default:
|
||||
// default regex for BPE tokenization pre-processing
|
||||
regex_exprs = {
|
||||
|
|
@ -2201,6 +2208,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|
|||
} else if (
|
||||
tokenizer_pre == "exaone4") {
|
||||
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
|
||||
} else if (
|
||||
tokenizer_pre == "exaone-moe") {
|
||||
pre_type = LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE;
|
||||
} else if (
|
||||
tokenizer_pre == "chameleon") {
|
||||
pre_type = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
|
||||
|
|
@ -2675,7 +2685,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|
|||
auto & attr = id_to_token[t.second].attr;
|
||||
|
||||
if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
|
||||
attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
|
||||
LLAMA_LOG_WARN("%s: setting token '%s' (%d) attribute to USER_DEFINED (%u), old attributes: %u\n",
|
||||
__func__, t.first.c_str(), t.second, LLAMA_TOKEN_ATTR_USER_DEFINED, attr);
|
||||
|
||||
attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -2728,7 +2741,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|
|||
special_eog_ids.erase(end_id);
|
||||
|
||||
auto & attr = id_to_token[end_id].attr;
|
||||
attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
|
||||
attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
|
||||
|
||||
LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>', or '<|calls|>' and '<|flush|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
|
||||
}
|
||||
|
|
@ -3565,34 +3578,34 @@ int32_t llama_vocab::impl::detokenize(
|
|||
}
|
||||
|
||||
void llama_vocab::impl::print_info() const {
|
||||
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, type_name().c_str());
|
||||
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, vocab.n_tokens());
|
||||
LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (uint32_t) bpe_ranks.size());
|
||||
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, type_name().c_str());
|
||||
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, vocab.n_tokens());
|
||||
LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (uint32_t) bpe_ranks.size());
|
||||
|
||||
// special tokens
|
||||
if (special_bos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, special_bos_id, id_to_token.at(special_bos_id).text.c_str() ); }
|
||||
if (special_eos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, special_eos_id, id_to_token.at(special_eos_id).text.c_str() ); }
|
||||
if (special_eot_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, special_eot_id, id_to_token.at(special_eot_id).text.c_str() ); }
|
||||
if (special_eom_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOM token = %d '%s'\n", __func__, special_eom_id, id_to_token.at(special_eom_id).text.c_str() ); }
|
||||
if (special_unk_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, special_unk_id, id_to_token.at(special_unk_id).text.c_str() ); }
|
||||
if (special_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, special_sep_id, id_to_token.at(special_sep_id).text.c_str() ); }
|
||||
if (special_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, special_pad_id, id_to_token.at(special_pad_id).text.c_str() ); }
|
||||
if (special_mask_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, special_mask_id, id_to_token.at(special_mask_id).text.c_str() ); }
|
||||
if (special_bos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, special_bos_id, id_to_token.at(special_bos_id).text.c_str() ); }
|
||||
if (special_eos_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, special_eos_id, id_to_token.at(special_eos_id).text.c_str() ); }
|
||||
if (special_eot_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, special_eot_id, id_to_token.at(special_eot_id).text.c_str() ); }
|
||||
if (special_eom_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: EOM token = %d '%s'\n", __func__, special_eom_id, id_to_token.at(special_eom_id).text.c_str() ); }
|
||||
if (special_unk_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, special_unk_id, id_to_token.at(special_unk_id).text.c_str() ); }
|
||||
if (special_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, special_sep_id, id_to_token.at(special_sep_id).text.c_str() ); }
|
||||
if (special_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, special_pad_id, id_to_token.at(special_pad_id).text.c_str() ); }
|
||||
if (special_mask_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, special_mask_id, id_to_token.at(special_mask_id).text.c_str() ); }
|
||||
|
||||
if (linefeed_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, linefeed_id, id_to_token.at(linefeed_id).text.c_str() ); }
|
||||
if (linefeed_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, linefeed_id, id_to_token.at(linefeed_id).text.c_str() ); }
|
||||
|
||||
if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
|
||||
if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
|
||||
if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
|
||||
if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
|
||||
if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
|
||||
if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
|
||||
if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
|
||||
if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
|
||||
if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
|
||||
if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
|
||||
if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
|
||||
if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
|
||||
|
||||
for (const auto & id : special_eog_ids) {
|
||||
LLAMA_LOG_INFO( "%s: EOG token = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
|
||||
LLAMA_LOG_INFO( "%s: EOG token = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
|
||||
}
|
||||
|
||||
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);
|
||||
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);
|
||||
}
|
||||
|
||||
llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
|
||||
|
|
|
|||
|
|
@ -54,6 +54,7 @@ enum llama_vocab_pre_type {
|
|||
LLAMA_VOCAB_PRE_TYPE_AFMOE = 42,
|
||||
LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43,
|
||||
LLAMA_VOCAB_PRE_TYPE_YOUTU = 44,
|
||||
LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE = 45,
|
||||
};
|
||||
|
||||
struct LLM_KV;
|
||||
|
|
|
|||
146
src/models/exaone-moe.cpp
Normal file
146
src/models/exaone-moe.cpp
Normal file
|
|
@ -0,0 +1,146 @@
|
|||
#include "models.h"
|
||||
|
||||
|
||||
llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
|
||||
llm_graph_context(params) {
|
||||
const int64_t n_embd_head = hparams.n_embd_head_k;
|
||||
|
||||
GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
|
||||
GGML_ASSERT(n_embd_head == hparams.n_rot);
|
||||
|
||||
ggml_tensor * cur;
|
||||
ggml_tensor * inpL;
|
||||
|
||||
inpL = build_inp_embd(model.tok_embd);
|
||||
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
auto * inp_attn_iswa = build_attn_inp_kv_iswa();
|
||||
|
||||
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||
|
||||
const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
|
||||
for (int il = 0; il < n_transformer_layers; ++il) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
// use RoPE for SWA layers
|
||||
const bool is_local_layer = hparams.is_swa(il);
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(cur, "attn_norm", il);
|
||||
|
||||
// self-attention
|
||||
{
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
cb(Qcur, "Qcur", il);
|
||||
|
||||
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
||||
|
||||
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
|
||||
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(Qcur, "Qcur_normed", il);
|
||||
cb(Kcur, "Kcur_normed", il);
|
||||
|
||||
if (is_local_layer) {
|
||||
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
|
||||
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
|
||||
Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
|
||||
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
}
|
||||
cb(Qcur, "Qcur", il);
|
||||
cb(Kcur, "Kcur", il);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
cur = build_attn(inp_attn_iswa,
|
||||
model.layers[il].wo, NULL,
|
||||
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
|
||||
cb(cur, "attn_out", il);
|
||||
}
|
||||
if (il == n_transformer_layers - 1 && inp_out_ids) {
|
||||
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
||||
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
||||
}
|
||||
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
// norm
|
||||
cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
// feed-forward network
|
||||
if (model.layers[il].ffn_gate_inp == nullptr) {
|
||||
// dense branch
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, NULL, NULL,
|
||||
model.layers[il].ffn_gate, NULL, NULL,
|
||||
model.layers[il].ffn_down, NULL, NULL, NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
// MoE branch
|
||||
ggml_tensor * moe_out = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
model.layers[il].ffn_up_exps,
|
||||
model.layers[il].ffn_gate_exps,
|
||||
model.layers[il].ffn_down_exps,
|
||||
model.layers[il].ffn_exp_probs_b,
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_SILU, hparams.expert_weights_norm,
|
||||
true, hparams.expert_weights_scale,
|
||||
(llama_expert_gating_func_type) hparams.expert_gating_func,
|
||||
il);
|
||||
cb(moe_out, "ffn_moe_out", il);
|
||||
|
||||
// FFN shared expert
|
||||
{
|
||||
ggml_tensor * ffn_shexp =
|
||||
build_ffn(cur,
|
||||
model.layers[il].ffn_up_shexp, NULL, NULL,
|
||||
model.layers[il].ffn_gate_shexp, NULL, NULL,
|
||||
model.layers[il].ffn_down_shexp, NULL, NULL,
|
||||
NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(ffn_shexp, "ffn_shexp", il);
|
||||
|
||||
cur = ggml_add(ctx0, moe_out, ffn_shexp);
|
||||
cb(cur, "ffn_out", il);
|
||||
}
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
|
||||
cur = build_cvec(cur, il);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
cur = inpL;
|
||||
|
||||
// final norm
|
||||
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
|
||||
|
||||
cb(cur, "result_norm", -1);
|
||||
res->t_embd = cur;
|
||||
|
||||
// lm_head
|
||||
cur = build_lora_mm(model.output, cur);
|
||||
|
||||
cb(cur, "result_output", -1);
|
||||
res->t_logits = cur;
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
}
|
||||
|
|
@ -167,6 +167,10 @@ struct llm_build_exaone : public llm_graph_context {
|
|||
llm_build_exaone(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
||||
struct llm_build_exaone_moe : public llm_graph_context {
|
||||
llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
||||
struct llm_build_falcon : public llm_graph_context {
|
||||
llm_build_falcon(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue