ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 11:21:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable CUDA graphs for MoE models + GPT-OSS support (#689)

Browse Source 
* gmp-oss: common

* gpt-oss: attnetion sinks, swiglu_oai

* gpt-oss: WIP llama

Model loads and runs (CPU only), but PPL is much to high
(~1500 for 1st batch vs ~200 in mainline).
Is it because of SWA, because of vocab, or did I introduce a bug somewhere?

* gpt-oss: CPU seems to be working

It was the SWA thta was missing in the previous commit.

There are issues with EOG tokens, so this still needs to be added.

* CUDA: ADD_ID

Just a copy from mainline

* gpt-oss: Seems to be working on CUDA

* gpt-oss: add sinks to the attn-vec kernels

* CUDA: add head size of 64 to new mma

Haven't turned it on yet, but observe slightly better PP and slightly
worse TG performance with that.

* gpt-oss: add ability to use -fmoe (only CUDA for now)

* Move row sums to the write place

* Add sinks to iqk flash attention

* gpt_oss: Implement -fmoe on the CPU

* Simdify swiglu_oai

Turning it off for now as performance becomes more variable,
so perhaps I'm running into thermal trottling imore often
because of making the CPU work too hard.

* llama: factor out model loader

* Builds successfully

* It runs, but mmap does not work

* Fix llama_mmap so mmap works

* Minor

* Fix CUDA after latest changes

* Attempt to use CUDA graphs with MoE models - not working

* CUDA graphs WIP - still not working

* CUDA graphs - seems to be working

Likely not all MLA variants are working.
I no longer remember why I added the q8_0 cpy that
transposes the tensor, but if really needed, this is now
missing. Also missing is q6_0.

* Make q8_0 cache work for DeepSeek models with CUDA graphs

* cuda: cpy for q6_0

* Fix llama_mmap on non-Linux platforms

* Adding forgotten file

* Iterating on Windows build failures

* cuda: re-add q8_0 -> q8_0 transpose

so mla = 2 can be used with CUDA graphs and q8_0 cache.

* Disable graphs without -fmoe

* Minor

* Turn graphs on by default

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-08-15 09:18:07 +03:00
committed by GitHub
parent c00335684c
commit 633e0617b0
56 changed files with 8720 additions and 5115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
common/chat.h
View File

@@ -13,20 +13,20 @@ struct common_chat_templates;
struct common_string_range {
size_t begin;
size_t end;
common_string_range(size_t begin, size_t end) : begin(begin), end(end) {
if (begin > end) {
throw std::runtime_error("Invalid range");
}
}
// prevent default ctor
common_string_range() = delete;
bool empty() const {
return begin == end;
}
bool operator==(const common_string_range & other) const {
return begin == other.begin && end == other.end;
}
@@ -40,7 +40,7 @@ struct common_chat_tool_call {
bool operator==(const common_chat_tool_call & other) const {
return name == other.name && arguments == other.arguments && id == other.id;
}
bool operator!=(const common_chat_tool_call & other) const {
return !(*this == other);
}
@@ -65,10 +65,10 @@ struct common_chat_msg {
std::string tool_call_id;
bool empty() const {
return content.empty() && content_parts.empty() && tool_calls.empty() &&
return content.empty() && content_parts.empty() && tool_calls.empty() &&
reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
}
void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
for (auto i = 0u; i < tool_calls.size(); i++) {
if (ids_cache.size() <= i) {
@@ -91,7 +91,7 @@ struct common_chat_msg {
&& tool_name == other.tool_name
&& tool_call_id == other.tool_call_id;
}
bool operator!=(const common_chat_msg & other) const {
return !(*this == other);
}
@@ -110,7 +110,7 @@ struct common_chat_msg_diff {
&& tool_call_index == other.tool_call_index
&& tool_call_delta == other.tool_call_delta;
}
bool operator!=(const common_chat_msg_diff & other) const {
return !(*this == other);
}
@@ -132,18 +132,20 @@ enum common_chat_format {
COMMON_CHAT_FORMAT_CONTENT_ONLY,
COMMON_CHAT_FORMAT_GENERIC,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
COMMON_CHAT_FORMAT_GPT_OSS,
COMMON_CHAT_FORMAT_KIMI_K2, // Our custom format (keep last for backward compatibility)
};
enum common_reasoning_format {
COMMON_REASONING_FORMAT_NONE,
COMMON_REASONING_FORMAT_AUTO,
COMMON_REASONING_FORMAT_DEEPSEEK,
COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY,
};
struct common_chat_syntax {
common_chat_format format = COMMON_CHAT_FORMAT_KIMI_K2;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO; //COMMON_REASONING_FORMAT_NONE;
// 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;
@@ -165,11 +167,12 @@ class common_chat_msg_partial_exception : public std::runtime_error {
// Format detection from chat template
common_chat_format common_chat_format_detect(const std::string & chat_template);
const char* common_chat_format_name(common_chat_format format);
const char* common_reasoning_format_name(common_reasoning_format format);
// Main parsing function (entry point for original llama.cpp compatibility)
common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
// Forward declare parser class
// Forward declare parser class
class common_chat_msg_parser;
// Format-specific parsing functions (accessible from chat-parser)