Fix new line issue with chat template, disable template when in-prefix/suffix is set (#8203)

* preserve new line llama_chat_format_single

* disable chat template if in-prefix/suffix is set

* remove redundant change

common/common.cpp

@@ -1014,16 +1014,19 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--in-prefix-bos") {
         params.input_prefix_bos = true;
+        params.enable_chat_template = false;
         return true;
     }
     if (arg == "--in-prefix") {
         CHECK_ARG
         params.input_prefix = argv[i];
+        params.enable_chat_template = false;
         return true;
     }
     if (arg == "--in-suffix") {
         CHECK_ARG
         params.input_suffix = argv[i];
+        params.enable_chat_template = false;
         return true;
     }
     if (arg == "--spm-infill") {
@@ -1406,7 +1409,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "halt generation at PROMPT, return control in interactive mode\n"
                                                                         "can be specified more than once for multiple prompts" });
     options.push_back({ "main",        "-sp,   --special",              "special tokens output enabled (default: %s)", params.special ? "true" : "false" });
-    options.push_back({ "main",        "-cnv,  --conversation",         "run in conversation mode (does not print special tokens and suffix/prefix) (default: %s)", params.conversation ? "true" : "false" });
+    options.push_back({ "main",        "-cnv,  --conversation",         "run in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: %s)", params.conversation ? "true" : "false" });
     options.push_back({ "main infill", "-i,    --interactive",          "run in interactive mode (default: %s)", params.interactive ? "true" : "false" });
     options.push_back({ "main infill", "-if,   --interactive-first",    "run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false" });
     options.push_back({ "main infill", "-mli,  --multiline-input",      "allows you to write or paste multiple lines without ending each in '\\'" });
@@ -2668,12 +2671,19 @@ std::string llama_chat_format_single(const struct llama_model * model,
         const std::vector<llama_chat_msg> & past_msg,
         const llama_chat_msg & new_msg,
         bool add_ass) {
+    std::ostringstream ss;
     auto fmt_past_msg = llama_chat_apply_template(model, tmpl, past_msg, false);
     std::vector<llama_chat_msg> chat_new(past_msg);
+    // if the past_msg ends with a newline, we must preserve it in the formatted version
+    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
+        ss << "\n";
+    };
+    // format chat with new_msg
     chat_new.push_back(new_msg);
     auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass);
-    auto formatted = fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
-    return formatted;
+    // get the diff part
+    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
+    return ss.str();
 }
 
 std::string llama_chat_format_example(const struct llama_model * model,

common/common.h

@@ -200,6 +200,7 @@ struct gpt_params {
     std::string public_path   = "";
     std::string chat_template = "";
     std::string system_prompt = "";
+    bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
 

convert-hf-to-gguf.py

@@ -2363,6 +2363,12 @@ class Gemma2Model(Model):
         self.gguf_writer.add_key_length(hparams["head_dim"])
         self.gguf_writer.add_value_length(hparams["head_dim"])
         self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_attn_logit_softcapping(
+            self.hparams["attn_logit_softcapping"]
+        )
+        self.gguf_writer.add_final_logit_softcapping(
+            self.hparams["final_logit_softcapping"]
+        )
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unusem

examples/main/main.cpp

@@ -261,7 +261,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd_inp;
 
     {
-        auto prompt = params.conversation
+        auto prompt = (params.conversation && params.enable_chat_template)
             ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
             : params.prompt;
         if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
@@ -810,7 +810,9 @@ int main(int argc, char ** argv) {
                         is_antiprompt = true;
                     }
 
-                    chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
+                    if (params.enable_chat_template) {
+                        chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
+                    }
                     is_interacting = true;
                     printf("\n");
                 }
@@ -872,12 +874,13 @@ int main(int argc, char ** argv) {
                         string_process_escapes(buffer);
                     }
 
-                    std::string user_inp = params.conversation
+                    bool format_chat = params.conversation && params.enable_chat_template;
+                    std::string user_inp = format_chat
                         ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
                         : std::move(buffer);
                     // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, user_inp,            false, params.conversation);
+                    const auto line_inp = ::llama_tokenize(ctx, user_inp,            false, format_chat);
                     const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
 
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());

gguf-py/gguf/constants.py

@@ -50,6 +50,8 @@ class Keys:
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
+        ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
+        FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"

gguf-py/gguf/gguf_writer.py

@@ -516,6 +516,12 @@ class GGUFWriter:
     def add_logit_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
 
+    def add_attn_logit_softcapping(self, value: float) -> None:
+        self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
+
+    def add_final_logit_softcapping(self, value: float) -> None:
+        self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
+
     def add_expert_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
 

src/llama.cpp

@@ -302,6 +302,8 @@ enum llm_kv {
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_FINAL_LOGIT_SOFTCAPPING,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -392,6 +394,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_POOLING_TYPE ,                     "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
+    { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
+    { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
@@ -2099,6 +2103,9 @@ struct llama_hparams {
     float f_norm_eps;
     float f_norm_rms_eps;
 
+    float f_attn_logit_softcapping = 50.0f;
+    float f_final_logit_softcapping = 30.0f;
+
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
     float    rope_freq_scale_train;
@@ -2115,8 +2122,9 @@ struct llama_hparams {
     float f_max_alibi_bias = 0.0f;
     float f_logit_scale    = 0.0f;
 
-    bool causal_attn = true;
-    bool use_alibi   = false;
+    bool causal_attn   = true;
+    bool use_alibi     = false;
+    bool attn_soft_cap = false;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
@@ -4702,6 +4710,9 @@ static void llm_load_hparams(
         case LLM_ARCH_GEMMA2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false);
+                ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
+                hparams.attn_soft_cap = true;
 
                 switch (hparams.n_layer) {
                     case 42: model.type = e_model::MODEL_9B; break;
@@ -7579,6 +7590,12 @@ static struct ggml_tensor * llm_build_kqv(
             kq = ggml_scale(ctx, kq, 30);
         }
 
+        if (hparams.attn_soft_cap) {
+            kq = ggml_scale(ctx, kq, 1.0f / hparams.f_attn_logit_softcapping);
+            kq = ggml_tanh(ctx, kq);
+            kq = ggml_scale(ctx, kq, hparams.f_attn_logit_softcapping);
+        }
+
         kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
         cb(kq, "kq_soft_max_ext", il);
 
@@ -11039,7 +11056,7 @@ struct llm_build_context {
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
 
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head)));
                 cb(Qcur, "Qcur_scaled", il);
 
                 Kcur = ggml_rope_ext(
@@ -11106,6 +11123,12 @@ struct llm_build_context {
 
         // lm_head
         cur = ggml_mul_mat(ctx0, model.output, cur);
+
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -17379,6 +17402,12 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
+    if (params.flash_attn && model->hparams.attn_soft_cap) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with attn_soft_cap - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
+
     if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
         LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
         params.flash_attn = false;

tests/test-chat-template.cpp

@@ -142,9 +142,9 @@ int main(void) {
         std::cout << "fmt_single(" << tmpl << ")\n" << output << "\n-------------------------\n";
         return output;
     };
-    assert(fmt_single("chatml") == "<|im_start|>user\nHow are you<|im_end|>\n<|im_start|>assistant\n");
+    assert(fmt_single("chatml") == "\n<|im_start|>user\nHow are you<|im_end|>\n<|im_start|>assistant\n");
     assert(fmt_single("llama2") == "[INST] How are you [/INST]");
-    assert(fmt_single("gemma") == "<start_of_turn>user\nHow are you<end_of_turn>\n<start_of_turn>model\n");
+    assert(fmt_single("gemma") == "\n<start_of_turn>user\nHow are you<end_of_turn>\n<start_of_turn>model\n");
     assert(fmt_single("llama3") == "<|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n");
 
     return 0;