tor-browser

commit 2187755546cc1548643844ed91ad9f64e8c4c4a1
parent def70ad6d72880f238f03d0994ab727ec11a3477
Author: Aristide Tossou <atossou@mozilla.com>
Date:   Wed,  8 Oct 2025 14:27:11 +0000

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

Differential Revision: https://phabricator.services.mozilla.com/D267614

Diffstat:
M
config/external/mozinference/mozinference.symbols
 | 
1
+
M
toolkit/components/ml/backends/llama/LlamaBackend.cpp
 | 
4
++++
M
toolkit/components/ml/backends/llama/LlamaRuntimeLinker.h
 | 
1
+

3 files changed, 6 insertions(+), 0 deletions(-)
diff --git a/config/external/mozinference/mozinference.symbols b/config/external/mozinference/mozinference.symbols
@@ -19,6 +19,7 @@ llama_sampler_chain_init
 llama_sampler_init_greedy
 llama_sampler_init_temp
 llama_sampler_init_dist
+llama_sampler_init_top_k
 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
@@ -376,6 +376,10 @@ LlamaBackend::SamplerResult LlamaBackend::InitializeSampler(
         break;
       }
 
+      case LlamaSamplerType::Top_k:
+        samplerElement = mLib->llama_sampler_init_top_k(samplerConfig.mTopK);
+        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
@@ -47,6 +47,7 @@ namespace mozilla::llama {
   X(struct llama_sampler*, llama_sampler_init_greedy, (void))               \
   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(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,                       \