remove --multi-token-probs

Author: ngxson

common/arg.cpp

@@ -1085,16 +1085,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
-    add_opt(common_arg(
-        {"-mtp", "--multi-token-probs"},
-        string_format(
-            "allow getting probabilities for multiple tokens. note: this will slow down the generation speed (default: %s)",
-            params.sampling.multi_token_probs ? "enabled" : "disabled"
-        ),
-        [](common_params & params) {
-            params.sampling.multi_token_probs = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MULTI_TOKEN_PROBS"));
     add_opt(common_arg(
         {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",

common/common.h

@@ -134,7 +134,6 @@ struct common_params_sampling {
     bool    ignore_eos         = false;
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
-    bool    multi_token_probs  = false; // output probabilities for multiple tokens (when n_probs > 0)
 
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 

examples/server/README.md

@@ -139,7 +139,6 @@ The project is under active development, and we are [looking for feedback and co
 | `-sp, --special` | special tokens output enabled (default: false) |
 | `--no-warmup` | skip warming up the model with an empty run |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
-| `-mtp, --multi-token-probs` | allow getting probabilities for multiple tokens. note: this will slow down the generation speed (default: disabled)<br/>(env: LLAMA_ARG_MULTI_TOKEN_PROBS) |
 | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |

examples/server/server.cpp

@@ -243,10 +243,6 @@ struct server_task {
         params.speculative.n_min = std::max(params.speculative.n_min, 2);
         params.speculative.n_max = std::max(params.speculative.n_max, 0);
 
-        if (!params_base.sampling.multi_token_probs && params.n_predict > 1 && params.sampling.n_probs > 0) {
-            throw std::runtime_error("For performance reason, n_probs with n_predict > 1 is not allowed. To enable this, start the server with --multi-token-probs");
-        }
-
         // TODO: add more sanity checks for the input parameters
 
         if (params.sampling.penalty_last_n < -1) {

examples/server/tests/unit/test_chat_completion.py

@@ -166,7 +166,6 @@ def test_chat_completion_with_timings_per_token():
 
 def test_logprobs():
     global server
-    server.multi_token_probs = True
     server.start()
     client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
     res = client.chat.completions.create(
@@ -194,7 +193,6 @@ def test_logprobs():
 
 def test_logprobs_stream():
     global server
-    server.multi_token_probs = True
     server.start()
     client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
     res = client.chat.completions.create(

examples/server/tests/unit/test_completion.py

@@ -259,7 +259,6 @@ def check_slots_status():
 
 def test_n_probs():
     global server
-    server.multi_token_probs = True
     server.start()
     res = server.make_request("POST", "/completion", data={
         "prompt": "I believe the meaning of life is",
@@ -285,7 +284,6 @@ def test_n_probs():
 
 def test_n_probs_stream():
     global server
-    server.multi_token_probs = True
     server.start()
     res = server.make_stream_request("POST", "/completion", data={
         "prompt": "I believe the meaning of life is",
@@ -313,7 +311,6 @@ def test_n_probs_stream():
 
 def test_n_probs_post_sampling():
     global server
-    server.multi_token_probs = True
     server.start()
     res = server.make_request("POST", "/completion", data={
         "prompt": "I believe the meaning of life is",

examples/server/tests/utils.py

@@ -74,7 +74,6 @@ class ServerProcess:
     draft_min: int | None = None
     draft_max: int | None = None
     no_webui: bool | None = None
-    multi_token_probs: bool | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -165,8 +164,6 @@ def start(self, timeout_seconds: int = 10) -> None:
             server_args.extend(["--draft-min", self.draft_min])
         if self.no_webui:
             server_args.append("--no-webui")
-        if self.multi_token_probs:
-            server_args.append("--multi-token-probs")
 
         args = [str(arg) for arg in [server_path, *server_args]]
         print(f"bench: starting server with: {' '.join(args)}")