tor-browser

commit b4db016217a94ccc1060c1876aa6531948ed811a
parent 2187755546cc1548643844ed91ad9f64e8c4c4a1
Author: Aristide Tossou <atossou@mozilla.com>
Date:   Wed,  8 Oct 2025 14:27:11 +0000

Bug 1990534 - Support TopP Sampling for llama.cpp. r=padenot

Differential Revision: https://phabricator.services.mozilla.com/D267615

Diffstat:
M
config/external/mozinference/mozinference.symbols
 | 
1
+
M
toolkit/components/ml/backends/llama/LlamaBackend.cpp
 | 
5
+++++
M
toolkit/components/ml/backends/llama/LlamaRuntimeLinker.h
 | 
2
++

3 files changed, 8 insertions(+), 0 deletions(-)
diff --git a/config/external/mozinference/mozinference.symbols b/config/external/mozinference/mozinference.symbols
@@ -20,6 +20,7 @@ llama_sampler_init_greedy
 llama_sampler_init_temp
 llama_sampler_init_dist
 llama_sampler_init_top_k
+llama_sampler_init_top_p
 llama_sampler_free
 llama_sampler_sample
 llama_memory_clear
diff --git a/toolkit/components/ml/backends/llama/LlamaBackend.cpp b/toolkit/components/ml/backends/llama/LlamaBackend.cpp
@@ -380,6 +380,11 @@ LlamaBackend::SamplerResult LlamaBackend::InitializeSampler(
         samplerElement = mLib->llama_sampler_init_top_k(samplerConfig.mTopK);
         break;
 
+      case LlamaSamplerType::Top_p:
+        samplerElement = mLib->llama_sampler_init_top_p(samplerConfig.mTopP,
+                                                        samplerConfig.mMinKeep);
+        break;
+
       default:
 
         auto msg = nsFmtCString(FMT_STRING("{}: Unimplemented sampler type"),
diff --git a/toolkit/components/ml/backends/llama/LlamaRuntimeLinker.h b/toolkit/components/ml/backends/llama/LlamaRuntimeLinker.h
@@ -48,6 +48,8 @@ namespace mozilla::llama {
   X(struct llama_sampler*, llama_sampler_init_temp, (float t))              \
   X(struct llama_sampler*, llama_sampler_init_dist, (uint32_t seed))        \
   X(struct llama_sampler*, llama_sampler_init_top_k, (int32_t k))           \
+  X(struct llama_sampler*, llama_sampler_init_top_p,                        \
+    (float p, size_t min_keep))                                             \
   X(void, llama_memory_clear, (llama_memory_t mem, bool data))              \
   X(llama_memory_t, llama_get_memory, (const struct llama_context* ctx))    \
   X(const struct llama_vocab*, llama_model_get_vocab,                       \