Merge branch 'master' into xsn/qwen3_embd_rerank

Author: ngxson

ci/run.sh

@@ -46,7 +46,20 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
+
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
+        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
+        else
+            echo "Warning: Using fallback CUDA architectures"
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
+        fi
+    else
+        echo "Error: nvidia-smi not found, cannot build with CUDA"
+        exit 1
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

convert_hf_to_gguf.py

@@ -3767,8 +3767,7 @@ def set_gguf_parameters(self):
         self._try_set_pooling_type()
 
         if self.cls_out_labels:
-            key_name = gguf.Keys.Classifier.OUTPUT_LABELS.format(arch = gguf.MODEL_ARCH_NAMES[self.model_arch])
-            self.gguf_writer.add_array(key_name, [v for k, v in sorted(self.cls_out_labels.items())])
+            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -196,6 +196,7 @@ enum vk_device_architecture {
     AMD_RDNA1,
     AMD_RDNA2,
     AMD_RDNA3,
+    INTEL_XE2,
 };
 
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
@@ -246,6 +247,34 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             }
             return vk_device_architecture::AMD_RDNA2;
         }
+    } else if (props.vendorID == VK_VENDOR_ID_INTEL) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool subgroup_size_control = false;
+
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                subgroup_size_control = true;
+            }
+        }
+
+        if (!subgroup_size_control) {
+            return vk_device_architecture::OTHER;
+        }
+
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
+        props2.pNext = &subgroup_size_control_props;
+        device.getProperties2(&props2);
+
+        if (subgroup_size_control_props.minSubgroupSize == 16) {
+            // Xe2 architecture uses SIMD16 while previous Xe and Gen architecture uses SIMD8.
+            // Minimum subgroup size matches the SIMD width so we distinguish architecture by checking this value.
+            // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
+            // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
+            return vk_device_architecture::INTEL_XE2;
+        }
     }
     return vk_device_architecture::OTHER;
 }
@@ -10263,8 +10292,9 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
     switch (props.vendorID) {
     case VK_VENDOR_ID_INTEL:
-        // Intel drivers don't support coopmat properly yet
-        return false;
+        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
+        // while some older hardware (ex. Arc A770) has performance regressions
+        return arch == vk_device_architecture::INTEL_XE2;
     case VK_VENDOR_ID_AMD:
         if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
             // Workaround for AMD proprietary driver reporting support on all GPUs

gguf-py/gguf/gguf_writer.py

@@ -935,6 +935,9 @@ def add_eot_token_id(self, id: int) -> None:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
+        self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
+
     # for vision models
 
     def add_clip_has_vision_encoder(self, value: bool) -> None:

include/llama.h

@@ -61,7 +61,10 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
-    struct llama_kv_cache;
+
+    typedef struct llama_memory_i * llama_memory_t;
+
+    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -493,9 +496,11 @@ extern "C" {
     DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
 
     LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
-    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
     LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
+    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
+
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
@@ -609,7 +614,78 @@ extern "C" {
                          int32_t   il_end);
 
     //
-    // KV cache
+    // Memory
+    //
+
+    // Clear the memory contents
+    LLAMA_API void llama_memory_clear(llama_memory_t mem);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API bool llama_memory_seq_rm(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_cp(
+            llama_memory_t mem,
+              llama_seq_id seq_id_src,
+              llama_seq_id seq_id_dst,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_memory_seq_keep(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_add(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                 llama_pos delta);
+
+    // Integer division of the positions by factor of `d > 1`
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_div(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                       int d);
+
+    // Returns the smallest position present in the memory for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_min(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Returns the largest position present in the memory for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_max(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Check if the memory supports shifting
+    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
+
+    //
+    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
     //
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
@@ -623,7 +699,7 @@ extern "C" {
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
     LLAMA_API void llama_kv_self_clear(
-            struct llama_context * ctx);
+                struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
@@ -694,22 +770,22 @@ extern "C" {
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    LLAMA_API DEPRECATED(void llama_kv_self_defrag(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
             "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
 
     // Check if the context supports KV cache shifting
     LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
 
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API DEPRECATED(void llama_kv_self_update(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
             "simply remove this call, updates are applied lazily on the next llama_decode()");
 
     //
     // State / sessions
     //
 
     // Returns the *actual* size in bytes of the state
-    // (logits, embedding and kv_cache)
+    // (logits, embedding and memory)
     // Only use when saving the state, not when restoring it, otherwise the size may be too small.
     LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
     LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -765,12 +841,12 @@ extern "C" {
                           size_t   n_token_count),
         "use llama_state_save_file instead");
 
-    // Get the exact size needed to copy the KV cache of a single sequence
+    // Get the exact size needed to copy the state of a single sequence
     LLAMA_API size_t llama_state_seq_get_size(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
-    // Copy the KV cache of a single sequence into the specified buffer
+    // Copy the state of a single sequence into the specified buffer
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
@@ -836,16 +912,16 @@ extern "C" {
     // For encode-decoder contexts, processes the batch using the encoder.
     // Can store the encoder output internally for later use by the decoder's cross-attention layers.
     //   0 - success
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // < 0 - error. the memory state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
     // Process a batch of tokens.
-    // Requires KV cache.
+    // Requires the context to have a memory.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the memory state is restored to the state before this call
     //    0 - success
     //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
     //    2 - aborted

src/CMakeLists.txt

@@ -20,7 +20,6 @@ add_library(llama
             llama-hparams.cpp
             llama-impl.cpp
             llama-io.cpp
-            llama-kv-cache.cpp
             llama-kv-cache-unified.cpp
             llama-kv-cache-unified-iswa.cpp
             llama-kv-cache-recurrent.cpp