QVAC-23752 feat[bc]: remove sliding-context support from the llm-addon

docs/website/content/docs/addons/llm-llamacpp/index.mdx

@@ -242,7 +242,6 @@ const config = {
 | frequency_penalty | float                                       | 0                            | Frequency penalty for sampling                        |
 | tools             | `"true"` or `"false"`                       | `"false"`                    | Enable tool calling with jinja templating             |
 | verbosity         | 0 – 3 (0=ERROR, 1=WARNING, 2=INFO, 3=DEBUG) | 0                            | Logging verbosity level                               |
-| n_discarded       | integer                                     | 0                            | Tokens to discard in sliding window context           |
 | main-gpu          | integer, `"integrated"`, or `"dedicated"`   | —                            | GPU selection for multi-GPU systems                   |
 
 

packages/llm-llamacpp/CMakeLists.txt

@@ -81,7 +81,6 @@ endif()
     ${PROJECT_SOURCE_DIR}/addon/src/handlers/LoadConfigHandlers.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/AsyncWeightsLoader.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/CacheManager.cpp
-    ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContextSlider.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContinuousBatchScheduler.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/GenerationParamsApply.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/LlamaLazyInitializeBackend.cpp
@@ -93,7 +92,6 @@ endif()
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/MtmdLlmContext.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/MultiRequestBatcher.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/TextLlmContext.cpp
-    ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContextShifter.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ReasoningBlockCompactor.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ModelMetadata.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/utils/LoggingMacros.cpp
@@ -141,7 +139,6 @@ if(BUILD_CLI)
     ${PROJECT_SOURCE_DIR}/addon/src/handlers/LoadConfigHandlers.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/AsyncWeightsLoader.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/CacheManager.cpp
-    ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContextSlider.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContinuousBatchScheduler.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/GenerationParamsApply.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/LlamaLazyInitializeBackend.cpp
@@ -153,7 +150,6 @@ if(BUILD_CLI)
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/MtmdLlmContext.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/MultiRequestBatcher.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/TextLlmContext.cpp
-    ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ContextShifter.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ReasoningBlockCompactor.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/model-interface/ModelMetadata.cpp
     ${PROJECT_SOURCE_DIR}/addon/src/utils/LoggingMacros.cpp

packages/llm-llamacpp/README.md

@@ -175,7 +175,6 @@ const config = {
 | frequency_penalty | float                                       | 0                            | Frequency penalty for sampling                        |
 | tools             | `"true"` or `"false"`                       | `"false"`                    | Enable tool calling with jinja templating             |
 | verbosity         | 0 – 3 (0=ERROR, 1=WARNING, 2=INFO, 3=DEBUG) | 0                            | Logging verbosity level                               |
-| n_discarded       | integer                                     | 0                            | Tokens to discard in sliding window context. In batch mode the sliding window is the per-sequence slot (`n_ctx / n_parallel`), so `n_discarded` is clamped to that per-slot window, not the full context; a value `>=` the slot cap is clamped and logs a warning |
 | main-gpu          | integer, `"integrated"`, or `"dedicated"`   | —                            | GPU selection for multi-GPU systems                   |
 | split-mode        | `"none"`, `"layer"`, or `"row"`             | `"none"`                     | How to split the model across GPUs ([details](./docs/multi-gpu.md)) |
 | tensor-split      | comma-separated proportions (e.g. `"1,1"`)  | —                            | GPU split ratios for layer/row parallelism ([details](./docs/multi-gpu.md)) |
@@ -190,7 +189,7 @@ const config = {
 The addon picks a safe KV-cache type when `cache-type-k`/`cache-type-v` are unset, and validates any explicit choice per backend:
 
 - **Auto-default:** on a **Metal / Vulkan GPU** (with flash attention on) both K and V default to **`q8_0`** — quality-neutral vs `f16` and ~47% smaller KV cache. **CPU** and **OpenCL (Adreno)** keep **`f16`** (ARM CPU `q8_0` has a quality/throughput cost; quantized KV is unsafe on OpenCL — see below). Finetuning manages its own KV types and is left untouched.
-- **OpenCL (Adreno) accepts only `f16`/`f32`/`bf16`:** any other cache type — quantized (`q8_0`, `q4_0`, `q4_1`, `q5_0`, …) or unrecognized — throws a `StatusError`. A quantized K or V cache aborts in `llama_kv_cache::update` on KV-cache shifts / cache management (sliding context, state restore) because ggml-opencl has no `F32→quantized` requantize kernel. Use `f16`/`f32`/`bf16`, or a Vulkan GPU / CPU.
+- **OpenCL (Adreno) accepts only `f16`/`f32`/`bf16`:** any other cache type — quantized (`q8_0`, `q4_0`, `q4_1`, `q5_0`, …) or unrecognized — throws a `StatusError`. A quantized K or V cache aborts in `llama_kv_cache::update` on cache management (reasoning-block compaction, state restore) because ggml-opencl has no `F32→quantized` requantize kernel. Use `f16`/`f32`/`bf16`, or a Vulkan GPU / CPU.
 - **Mixed K≠V is a warning, not an error:** if K and V differ and at least one is quantized, the addon logs a warning (asymmetric quantized K/V falls off the fused flash-attention path — a notable GPU decode penalty — for no quality benefit, and is unsupported on Adreno OpenCL) but proceeds. Prefer a symmetric type. (This may be relaxed once qvac-fabric handles asymmetric quantized K/V efficiently.)
 
 

packages/llm-llamacpp/addon/src/addon/LlmErrors.hpp

@@ -31,6 +31,8 @@ enum LlmErrorCode : uint32_t {
   UnableToLoadMetadata = 23,
   ReloadNotSupportedForStreamedModel = 24,
   UnableToSaveSessionFile = 25,
+  // Retired with sliding-context support and never thrown now. Kept so a new
+  // code does not reuse 26, which builds up to 0.45.0 still emit.
   ContextSlideFailed = 26,
   Cancelled = 27,
   // mode llm spesific errors here

packages/llm-llamacpp/addon/src/model-interface/CacheManager.cpp

@@ -1,6 +1,5 @@
 #include "CacheManager.hpp"
 
-#include <array>
 #include <filesystem>
 #include <system_error>
 
@@ -19,58 +18,9 @@ using namespace qvac_lib_inference_addon_llama::errors;
 using namespace qvac_lib_inference_addon_cpp::logger;
 using namespace qvac_lib_inference_addon_llama::logging;
 
-namespace {
-
-struct SessionMetadata {
-  std::array<llama_token, SESSION_METADATA_FIELD_COUNT> tokens = {};
-
-  static SessionMetadata fromContext(const LlmContext& context) {
-    SessionMetadata metadata;
-    auto& tokens = metadata.tokens;
-    using Field = SessionMetadataField;
-    tokens[static_cast<size_t>(Field::NPast)] =
-        static_cast<llama_token>(context.getNPast());
-    tokens[static_cast<size_t>(Field::FirstMsgTokens)] =
-        static_cast<llama_token>(context.getFirstMsgTokens());
-    tokens[static_cast<size_t>(Field::CacheTokens)] =
-        static_cast<llama_token>(context.getCacheTokens());
-    tokens[static_cast<size_t>(Field::FirstMsgCacheTokens)] =
-        static_cast<llama_token>(context.getFirstMsgCacheTokens());
-    return metadata;
-  }
-
-  llama_token* data() { return tokens.data(); }
-  const llama_token* data() const { return tokens.data(); }
-  size_t size() const { return tokens.size(); }
-
-  llama_token field(SessionMetadataField which) const {
-    return tokens[static_cast<size_t>(which)];
-  }
-  llama_token nPast() const { return field(SessionMetadataField::NPast); }
-  llama_token firstMsgTokens() const {
-    return field(SessionMetadataField::FirstMsgTokens);
-  }
-  llama_token cacheTokens() const {
-    return field(SessionMetadataField::CacheTokens);
-  }
-  llama_token firstMsgCacheTokens() const {
-    return field(SessionMetadataField::FirstMsgCacheTokens);
-  }
-
-  void applyTo(LlmContext& context) const {
-    context.setNPast(nPast());
-    context.setFirstMsgTokens(firstMsgTokens());
-    context.setCacheTokens(cacheTokens());
-    context.setFirstMsgCacheTokens(firstMsgCacheTokens());
-  }
-};
-
-} // namespace
-
 CacheManager::CacheManager(
-    LlmContext* llmContext, llama_pos configuredNDiscarded,
-    std::function<void(bool)> resetStateCallback)
-    : llmContext_(llmContext), configuredNDiscarded_(configuredNDiscarded),
+    LlmContext* llmContext, std::function<void(bool)> resetStateCallback)
+    : llmContext_(llmContext),
       resetStateCallback_(std::move(resetStateCallback)) {}
 
 bool CacheManager::isFileInitialized(const std::filesystem::path& path) {
@@ -255,14 +205,6 @@ bool CacheManager::loadCache() {
   }
   sessionMetadata.applyTo(*llmContext_);
 
-  if (configuredNDiscarded_ >
-      llama_n_ctx(ctx) - llmContext_->getFirstMsgTokens()) {
-    llmContext_->setNDiscarded(
-        llama_n_ctx(ctx) - llmContext_->getFirstMsgTokens() - 1);
-  } else {
-    llmContext_->setNDiscarded(configuredNDiscarded_);
-  }
-
   auto* mem = llama_get_memory(ctx);
   if (mem == nullptr) {
     throw qvac_errors::StatusError(
@@ -371,7 +313,7 @@ void CacheManager::writeCacheFile(const std::string& path) {
       Priority::DEBUG,
       string_format("%s: saving cache to '%s'\n", __func__, path.c_str()));
   const SessionMetadata sessionMetadata =
-      SessionMetadata::fromContext(*llmContext_);
+      SessionMetadata::capture(*llmContext_);
   if (llama_state_seq_save_file(
           ctx,
           tmpPath.c_str(),

packages/llm-llamacpp/addon/src/model-interface/CacheManager.hpp

@@ -29,8 +29,7 @@ struct ParsedPromptPayload {
 class CacheManager {
 public:
   CacheManager(
-      LlmContext* llmContext, llama_pos configuredNDiscarded,
-      std::function<void(bool)> resetStateCallback);
+      LlmContext* llmContext, std::function<void(bool)> resetStateCallback);
 
   bool handleCache(
       ParsedPromptPayload& parsedPrompt, const std::string& inputPrompt,
@@ -54,7 +53,6 @@ class CacheManager {
   static bool isParentDirectoryMissing(const std::filesystem::path& path);
 
   LlmContext* llmContext_;
-  llama_pos configuredNDiscarded_;
   std::function<void(bool)> resetStateCallback_;
   std::string sessionPath_;
   bool cacheDisabled_ = true;