SYCL: Add non contiguous support in RMS_NORM and NORM kernels (#13611)

* SYCL: Add non contiguous input support to norm kernel

* refactor and add RMS_NORM non contiguous input support

ggml-ci

* restore subgroup reduction for multi-subgroup thread blocks in norm kernels

* Swap grid dims of nsamples and nrows

ggml-ci

* Revert "Swap grid dims of nsamples and nrows"

This reverts commit 43be2d657fec7f7fba54e2cd154106bc0fc45adf.

* restore not required changes
ggml-ci

* address review comments: change it to more like SYCL

* Use a common function to calculate offset

* remove wrap around logic for handling broadcasts

* remove static from calculate_offset fn and use ceil_div

common/chat-parser.cpp

@@ -170,20 +170,23 @@ std::string common_chat_msg_parser::consume_rest() {
 }
 
 // Tries to find the regex, consumes it (pos right after it) and gives the prelude (right before it) and the groups to the callback.
-std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from) {
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from, bool add_prelude_to_content) {
     auto m = regex.search(input_, from == std::string::npos ? pos_ : from);
     if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
         return std::nullopt;
     }
+    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
+    pos_ = m.groups[0].end;
+
+    if (add_prelude_to_content) {
+        add_content(prelude);
+    }
     if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
         if (is_partial()) {
             throw common_chat_msg_partial_exception(regex.str());
         }
         return std::nullopt;
     }
-    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
-    pos_ = m.groups[0].end;
-
     return find_regex_result{prelude, m.groups};
 }
 

common/chat-parser.h

@@ -30,6 +30,7 @@ class common_chat_msg_parser {
     const std::string & healing_marker() const { return healing_marker_; }
     const bool & is_partial() const { return is_partial_; }
     const common_chat_msg & result() const { return result_; }
+    const common_chat_syntax & syntax() const { return syntax_; }
 
     void move_to(size_t pos) {
         if (pos > input_.size()) {
@@ -77,7 +78,7 @@ class common_chat_msg_parser {
         std::vector<common_string_range> groups;
     };
 
-    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos);
+    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
 
     bool try_consume_literal(const std::string & literal);
 

common/chat.cpp

@@ -656,7 +656,6 @@ static void parse_json_tool_calls(
                 }
                 from = std::string::npos;
 
-                builder.add_content(res->prelude);
                 auto maybe_raw_python = name == "python" && allow_raw_python;
                 if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) {
                     if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) {
@@ -686,7 +685,6 @@ static void parse_json_tool_calls(
     };
     if (block_open) {
         if (auto res = builder.try_find_regex(*block_open)) {
-            builder.add_content(res->prelude);
             parse_tool_calls();
         } else {
             builder.add_content(builder.consume_rest());
@@ -699,7 +697,6 @@ static void parse_json_tool_calls(
 static void parse_prefixed_json_tool_call_array(common_chat_msg_parser & builder, const common_regex & prefix, size_t rstrip_prefix = 0) {
     static const std::vector<std::vector<std::string>> args_paths = {{"arguments"}};
     if (auto res = builder.try_find_regex(prefix)) {
-        builder.add_content(res->prelude);
         builder.move_back(rstrip_prefix);
         auto tool_calls = builder.consume_json_with_dumped_args(args_paths);
         if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) {
@@ -835,6 +832,10 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
     return data;
 }
 static void common_chat_parse_generic(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
     static const std::vector<std::vector<std::string>> content_paths = {
         {"response"},
     };
@@ -907,6 +908,11 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     return data;
 }
 static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
     static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
     parse_prefixed_json_tool_call_array(builder, prefix);
 }
@@ -1001,7 +1007,6 @@ static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) {
 
     if (auto res = builder.try_find_regex(start_action_regex)) {
         // If we didn't extract thoughts, prelude includes them.
-        builder.add_content(res->prelude);
         auto tool_calls = builder.consume_json_with_dumped_args({{"parameters"}});
         for (const auto & tool_call : tool_calls.value) {
             std::string name = tool_call.contains("tool_name") ? tool_call.at("tool_name") : "";
@@ -1016,11 +1021,7 @@ static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) {
         }
         builder.consume_regex(end_action_regex);
     } else if (auto res = builder.try_find_regex(start_response_regex)) {
-        // If we didn't extract thoughts, prelude includes them.
-        builder.add_content(res->prelude);
-        if (auto res = builder.try_find_regex(end_response_regex)) {
-            builder.add_content(res->prelude);
-        } else {
+        if (!builder.try_find_regex(end_response_regex)) {
             builder.add_content(builder.consume_rest());
             throw common_chat_msg_partial_exception(end_response_regex.str());
         }
@@ -1128,6 +1129,11 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     return data;
 }
 static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
     static const common_regex function_regex(
         "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: ");
     static const common_regex close_regex("\\}\\s*");
@@ -1138,8 +1144,6 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
     if (with_builtin_tools) {
         static const common_regex builtin_call_regex("<\\|python_tag\\|>");
         if (auto res = builder.try_find_regex(builtin_call_regex)) {
-            builder.add_content(res->prelude);
-
             auto fun_res = builder.consume_regex(function_name_regex);
             auto function_name = builder.str(fun_res.groups[1]);
 
@@ -1255,6 +1259,10 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
 }
 static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
     builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
 
     static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
     static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
@@ -1316,6 +1324,10 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
     return data;
 }
 static void common_chat_parse_firefunction_v2(common_chat_msg_parser & builder) {
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
     static const common_regex prefix(regex_escape(" functools["));
     parse_prefixed_json_tool_call_array(builder, prefix, /* rstrip_prefix= */ 1);
 }
@@ -1457,15 +1469,12 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
     return data;
 }
 static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser & builder) {
-    // This version of Functionary still supports the llama 3.1 tool call format for the python tool.
-    static const common_regex python_tag_regex(regex_escape("<|python_tag|>"));
-
-    if (auto res = builder.try_find_regex(python_tag_regex)) {
-        builder.add_content(res->prelude);
-        auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
-        builder.add_tool_call("python", "", arguments);
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
         return;
     }
+    // This version of Functionary still supports the llama 3.1 tool call format for the python tool.
+    static const common_regex python_tag_regex(regex_escape("<|python_tag|>"));
 
     static const common_regex function_regex(R"(<function=(\w+)>)");
     static const common_regex close_regex(R"(</function>)");
@@ -1477,6 +1486,12 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
         function_regex,
         close_regex,
         std::nullopt);
+
+    if (auto res = builder.try_find_regex(python_tag_regex)) {
+        auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
+        builder.add_tool_call("python", "", arguments);
+        return;
+    }
 }
 
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
@@ -1595,6 +1610,10 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
 }
 static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
     builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
 
     static const common_regex open_regex(
         "(?:"
@@ -1616,8 +1635,6 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
     );
 
     if (auto res = builder.try_find_regex(open_regex)) {
-        builder.add_content(res->prelude);
-
         const auto & block_start = res->groups[1];
         std::string block_end = block_start.empty() ? "" : "```";
 
@@ -1853,10 +1870,10 @@ static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
     builder.add_content(builder.consume_rest());
 }
 
-static void common_chat_parse(common_chat_msg_parser & builder, common_chat_format format) {
-    LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(format), builder.input().c_str());
+static void common_chat_parse(common_chat_msg_parser & builder) {
+    LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(builder.syntax().format), builder.input().c_str());
 
-    switch (format) {
+    switch (builder.syntax().format) {
         case COMMON_CHAT_FORMAT_CONTENT_ONLY:
             common_chat_parse_content_only(builder);
             break;
@@ -1891,7 +1908,7 @@ static void common_chat_parse(common_chat_msg_parser & builder, common_chat_form
             common_chat_parse_command_r7b(builder);
             break;
         default:
-            throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(format));
+            throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }
     builder.finish();
 }
@@ -1899,7 +1916,7 @@ static void common_chat_parse(common_chat_msg_parser & builder, common_chat_form
 common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
     common_chat_msg_parser builder(input, is_partial, syntax);
     try {
-        common_chat_parse(builder, syntax.format);
+        common_chat_parse(builder);
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {

common/chat.h

@@ -144,6 +144,7 @@ struct common_chat_syntax {
     // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
     bool                     reasoning_in_content  = false;
     bool                     thinking_forced_open  = false;
+    bool                     parse_tool_calls      = true;
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid

examples/training/README.md

@@ -10,8 +10,8 @@ Proof of concept:
 
 ``` sh
 export model_name=llama_3.2-1b && export quantization=f32
-./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
-./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+./build/bin/llama-finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/llama-perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
 ```
 
 The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

ggml/src/ggml-sycl/common.hpp

@@ -13,6 +13,7 @@
 #ifndef GGML_SYCL_COMMON_HPP
 #define GGML_SYCL_COMMON_HPP
 
+#include <cstddef>
 #include <fstream>
 #include <iostream>
 #include <string>
@@ -481,6 +482,19 @@ static __dpct_inline__ float warp_reduce_max(float x,
     return x;
 }
 
+/* Helper for Computing the linear offset of a ggml_tensor given
+per-dimension sizes, strides, and indices */
+template<int N>
+__dpct_inline__ size_t calculate_offset(const std::array<int, N> & strides, const std::array<int, N> & indices) {
+    size_t offset = 0;
+#pragma unroll
+    for (int i = 0; i < N; i++) {
+        auto index_i = indices[i];
+        offset += strides[i] * index_i;
+    }
+    return offset;
+}
+
 // Helper for vec loading aligned data
 template <typename Tp, int n>
 inline sycl::vec<Tp, n> vec_aligned_load(const Tp* aligned_ptr) {

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4241,6 +4241,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
 #endif
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
+            return true;
         case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);

ggml/src/ggml-sycl/norm.cpp

@@ -1,40 +1,50 @@
 #include "norm.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
 
-static void norm_f32(const float* x, float* dst, const int ncols, const float eps,
-    const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
-    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
-        item_ct1.get_local_id(1);
-    const int tid = item_ct1.get_local_id(2);
+static void norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
+        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
+
+    const int nrows = item_ct1.get_group_range(2);
+    const int nchannels = item_ct1.get_group_range(1);
 
     const int nthreads = item_ct1.get_local_range(2);
+    const int sample  = item_ct1.get_group(0);
+    const int channel = item_ct1.get_group(1);
+    const int row     = item_ct1.get_group(2);
+
+    const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
+
+    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
+    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+
+    x += strided_offset;
+    dst += packed_offset;
+
     sycl::float2 mean_var = sycl::float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row * ncols + col];
+        const float xi = x[col];
         mean_var.x() += xi;
         mean_var.y() += xi * xi;
     }
 
     // sum up partial sums
     mean_var = warp_reduce_sum(mean_var, item_ct1);
-    if (block_size > WARP_SIZE) {
-
-        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = mean_var;
+    if  (block_size > WARP_SIZE) {
+        const auto sub_group = item_ct1.get_sub_group();
+        const auto sg_id = sub_group.get_group_linear_id();
+        const auto wi_in_sg = sub_group.get_local_linear_id();
+        if (wi_in_sg == 0) {
+            s_sum[sg_id] = mean_var;
         }
-        /*
-        DPCT1118:0: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
         item_ct1.barrier(sycl::access::fence_space::local_space);
         mean_var = 0.f;
-        size_t nreduce = nwarps / WARP_SIZE;
+        const size_t nreduce = ceil_div(nwarps, WARP_SIZE);
         for (size_t i = 0; i < nreduce; i += 1)
         {
-            mean_var += s_sum[lane_id + i * WARP_SIZE];
+            mean_var += s_sum[wi_in_sg + i * WARP_SIZE];
         }
         mean_var = warp_reduce_sum(mean_var, item_ct1);
     }
@@ -44,7 +54,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
     const float inv_std = sycl::rsqrt(var + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row * ncols + col] = (x[row * ncols + col] - mean) * inv_std;
+        dst[col] = (x[col] - mean) * inv_std;
     }
 }
 
@@ -135,39 +145,51 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     }
 }
 
-static void rms_norm_f32(const float* x, float* dst, const int ncols, const float eps,
-    const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
-    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
-        item_ct1.get_local_id(1);
-    const int tid = item_ct1.get_local_id(2);
+static void rms_norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
+        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+
+    const int nrows = item_ct1.get_group_range(2);
+    const int nchannels = item_ct1.get_group_range(1);
+
+    const int sample  = item_ct1.get_group(0);
+    const int channel = item_ct1.get_group(1);
+    const int row     = item_ct1.get_group(2);
+
     const int nthreads = item_ct1.get_local_range(2);
+
+    const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
+
+    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
+    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+
+    x   += strided_offset;
+    dst += packed_offset;
+
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row * ncols + col];
+        const float xi = x[col];
         tmp += xi * xi;
     }
 
     // sum up partial sums
     tmp = warp_reduce_sum(tmp, item_ct1);
     if (block_size > WARP_SIZE) {
-
-        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
+        const auto sub_group = item_ct1.get_sub_group();
+        const auto sg_id = sub_group.get_group_linear_id();
+        const auto wi_in_sg = sub_group.get_local_linear_id();
+        if (wi_in_sg == 0) {
+            s_sum[sg_id] = tmp;
         }
-        /*
-        DPCT1118:3: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
+
         item_ct1.barrier(sycl::access::fence_space::local_space);
-        size_t nreduce = nwarps / WARP_SIZE;
+        const size_t nreduce = ceil_div(nwarps, WARP_SIZE);
         tmp = 0.f;
         for (size_t i = 0; i < nreduce; i += 1)
         {
-            tmp += s_sum[lane_id + i * WARP_SIZE];
+            tmp += s_sum[wi_in_sg + i * WARP_SIZE];
         }
         tmp = warp_reduce_sum(tmp, item_ct1);
     }
@@ -176,7 +198,7 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
     const float scale = sycl::rsqrt(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row * ncols + col] = scale * x[row * ncols + col];
+        dst[col] = scale * x[col];
     }
 }
 
@@ -224,20 +246,20 @@ static void l2_norm_f32(const float* x, float* dst, const int ncols, const float
     }
 }
 
-static void norm_f32_sycl(const float* x, float* dst, const int ncols,
-    const int nrows, const float eps,
-    queue_ptr stream, int device) {
+static void norm_f32_sycl(const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample,
+        const float eps, queue_ptr stream, int device) {
+
+    const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -252,15 +274,12 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
         */
         stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
-                sycl::range<1>(work_group_size / WARP_SIZE), cgh);
-
+                            sycl::range<1>(work_group_size / WARP_SIZE), cgh);
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -313,21 +332,20 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     }
 }
 
-static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
-    const int nrows, const float eps,
-    queue_ptr stream, int device) {
+static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
+
+    const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -344,12 +362,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -398,12 +414,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
 }
 
 void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+    const ggml_tensor * src0 = dst->src[0];
 
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows = ggml_nrows(dst->src[0]);
+    GGML_TENSOR_UNARY_OP_LOCALS
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
@@ -411,8 +427,14 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
 
-    norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
@@ -436,11 +458,10 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
+    const ggml_tensor * src0 = dst->src[0];
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows = ggml_nrows(dst->src[0]);
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
 
@@ -450,7 +471,13 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+    GGML_TENSOR_UNARY_OP_LOCALS
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
+    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {

tests/test-chat.cpp

@@ -401,9 +401,12 @@ static common_chat_msg simple_assist_msg(const std::string & content, const std:
     }
     return msg;
 }
-const common_chat_msg message_assist                             = simple_assist_msg("Hello, world!\nWhat's up?");
-const common_chat_msg message_assist_empty                       = simple_assist_msg("");
-const common_chat_msg message_assist_thoughts_unparsed_deepseek  = simple_assist_msg("<think>I'm\nthinking</think>Hello, world!\nWhat's up?");
+const common_chat_msg message_assist                              = simple_assist_msg("Hello, world!\nWhat's up?");
+const common_chat_msg message_assist_empty                        = simple_assist_msg("");
+const common_chat_msg message_assist_thoughts_unparsed_deepseek   = simple_assist_msg("<think>I'm\nthinking</think>Hello, world!\nWhat's up?");
+const common_chat_msg message_assist_thoughts_unparsed_md         = simple_assist_msg("<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}```");
+const common_chat_msg message_assist_thoughts_unparsed_md_partial = simple_assist_msg("<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}");
+
 const common_chat_msg message_assist_thoughts_unparsed_r7b       = simple_assist_msg("<|START_THINKING|>I'm\nthinking<|END_THINKING|>Hello, world!\nWhat's up?");
 const common_chat_msg message_assist_thoughts                    = simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking");
 const common_chat_msg message_assist_thoughts_unopened_unparsed  = simple_assist_msg("I'm\nthinking</think>Hello, world!\nWhat's up?");
@@ -591,8 +594,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_COMMAND_R7B,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
             common_chat_parse(
@@ -619,8 +620,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_COMMAND_R7B,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_call_idx,
             common_chat_parse(
@@ -632,8 +631,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_COMMAND_R7B,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_no_content,
             common_chat_parse(
@@ -644,8 +641,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_COMMAND_R7B,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
 
         test_templates(tmpls.get(), end_tokens, message_assist_call_idx, tools,
@@ -675,6 +670,18 @@ static void test_template_output_parsers() {
 
         // Generic tool calls doesn't generate / parse content-only messages symmetrically.
 
+        assert_equals(
+            simple_assist_msg("{ \"tool_call\" : { \"name\" : \"t"),
+            common_chat_parse(
+                "{ \"tool_call\" : { \"name\" : \"t",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GENERIC,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+                    /* .reasoning_in_content = */ false,
+                    /* .thinking_forced_open = */ true,
+                    /* .parse_tool_calls = */ false,
+                }));
         assert_equals(
             message_assist_empty,
             common_chat_parse(
@@ -776,11 +783,9 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(
-            simple_assist_msg(""),
+            simple_assist_msg("Let's call something\n"),
             common_chat_parse(
                 "Let's call something\n"
                 "<tool_call>{\"name",
@@ -788,8 +793,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_call_thoughts,
             common_chat_parse(
@@ -979,7 +982,34 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
+                }));
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse(
+                "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+                }));
+        assert_msg_equals(message_assist_thoughts_unparsed_md,
+            common_chat_parse(
+                "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}```",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+                    /* .reasoning_in_content = */ true,
+                    /* .thinking_forced_open = */ false,
+                    /* .parse_tool_calls = */ false,
+                }));
+        assert_msg_equals(message_assist_thoughts_unparsed_md_partial,
+            common_chat_parse(
+                "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}```",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+                    /* .reasoning_in_content = */ true,
                     /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unopened_unparsed,
@@ -989,8 +1019,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_HERMES_2_PRO,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts,
             common_chat_parse(
@@ -1073,6 +1101,13 @@ static void test_template_output_parsers() {
                     {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1}));
         }
 
+        assert_equals(
+            message_assist_call,
+            common_chat_parse(
+                "<function=special_function>{\"arg1\": 1}<",
+                /* is_partial= */ true,
+                {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1}));
+
         test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "<function=special_function>{\"arg1\": 1}</function>");
@@ -1187,8 +1222,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_R1,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unopened_unparsed,
             common_chat_parse(
@@ -1197,8 +1230,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_R1,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts,
             common_chat_parse(
@@ -1252,8 +1283,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_R1,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts,
             common_chat_parse(
@@ -1295,8 +1324,6 @@ static void test_template_output_parsers() {
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_R1,
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-                    /* .reasoning_in_content = */ false,
-                    /* .thinking_forced_open = */ false,
                 }));
         test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                 "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"

tools/server/server.cpp

@@ -364,6 +364,7 @@ struct server_task {
             params.oaicompat_chat_syntax.reasoning_format = params_base.reasoning_format;
             params.oaicompat_chat_syntax.reasoning_in_content = params.stream;
             params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
+            params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
         }
 
         {

tools/server/utils.hpp

@@ -735,8 +735,11 @@ static json oaicompat_chat_params_parse(
     inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
     inputs.reasoning_format      = opt.reasoning_format;
     inputs.enable_thinking       = opt.enable_thinking;
-    if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
-        throw std::runtime_error("Cannot use custom grammar constraints with tools.");
+    if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+        if (body.contains("grammar")) {
+            throw std::runtime_error("Cannot use custom grammar constraints with tools.");
+        }
+        llama_params["parse_tool_calls"] = true;
     }
 
     // if the assistant message appears at the end of list, we do not add end-of-turn token