server: save only restorable slot states to the prompt cache (#30498)
* server: save only restorable slot states to the prompt cache The idle-slot save loop stores every idle slot whenever a task starts, while only completion tasks restore from the prompt cache. Rerank, embedding, decision and infill states fill the cache up to its limit without ever being read back, and slots already cached are offered again at every task start. Each slot now tracks with cache_pending whether its state is restorable and missing from the cache: a completion release or an explicit slot restore sets it, prompt_save consumes it, prompt_clear resets it. Both save paths share this rule, and the LRU path restores only for completion tasks. * server: test that decision slots skip the prompt cache Decision tasks never mark their slot as cacheable, so starting new tasks must not save their idle slots to the RAM prompt cache, with or without a unified KV cache. * server: restore any reusing task from the prompt cache Embedding and rerank tasks never reuse their state, so their slots are cleared on release like child slots and the cache only holds states that a later task can restore. Infill and decision tasks now restore from the cache too, which replaces the cache_pending flag. Decision models disable the prompt cache unless --cache-ram is given, since their states seldom repeat.
This commit is contained in:
1 parent
843e7dcc15
commit
2639a3b4e0
8 files changed
+73
-9
No files matched your search
+2
-1
@@ -1717,10 +1717,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
|
||||
).set_env("LLAMA_ARG_CHECKPOINT_MIN_SPACING_NT").set_examples({LLAMA_EXAMPLE_SERVER}));
|
||||
add_opt(common_arg(
|
||||
{"-cram", "--cache-ram"}, "N",
|
||||
string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)"
|
||||
string_format("set the maximum cache size in MiB (default: %d, 0 for decision models, -1 - no limit, 0 - disable)"
|
||||
"[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)", params.cache_ram_mib),
|
||||
[](common_params & params, int value) {
|
||||
params.cache_ram_mib = value;
|
||||
params.cache_ram_explicit = true;
|
||||
}
|
||||
).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
|
||||
add_opt(common_arg(
|
||||
|
||||
@@ -640,6 +640,7 @@ struct common_params {
|
||||
int32_t kv_unified_per_slot = 0; // max context per parallel slot; 0 = unset
|
||||
int32_t checkpoint_min_step = 8192; // minimum spacing between context checkpoints
|
||||
int32_t cache_ram_mib = 8192; // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
|
||||
bool cache_ram_explicit = false; // set by --cache-ram, decision models disable the cache otherwise
|
||||
|
||||
std::string public_path = ""; // NOLINT
|
||||
std::string api_prefix = ""; // NOLINT
|
||||
|
||||
+1
-1
@@ -148,7 +148,7 @@
|
||||
| `--display-prompt, --no-display-prompt` | whether to print prompt at generation (default: true) |
|
||||
| `-co, --color [on\|off\|auto]` | Colorize output to distinguish prompt and user input from generations ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal |
|
||||
| `-ctxcp, --ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 32)[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
|
||||
| `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
|
||||
| `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, 0 for decision models, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
|
||||
| `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
|
||||
| `-sys, --system-prompt PROMPT` | system prompt to use with model (if applicable, depending on chat template) |
|
||||
| `--show-timings, --no-show-timings` | whether to show timing information after each response (default: true)<br/>(env: LLAMA_ARG_SHOW_TIMINGS) |
|
||||
|
||||
@@ -165,7 +165,7 @@ For the full list of features, please refer to [server's changelog](https://gith
|
||||
| `--kv-unified-per-slot N` | context limit per parallel slot (default: unset, behavior unchanged).<br/>when set without -c/--ctx-size, the shared KV pool is sized to n_parallel*N<br/>(env: LLAMA_ARG_KV_UNIFIED_PER_SLOT) |
|
||||
| `-ctxcp, --ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 32)[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
|
||||
| `-cms, --checkpoint-min-step N` | minimum spacing between context checkpoints in tokens (default: 8192, 0 = no minimum)<br/>(env: LLAMA_ARG_CHECKPOINT_MIN_SPACING_NT) |
|
||||
| `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
|
||||
| `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, 0 for decision models, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
|
||||
| `-kvu, --kv-unified, -no-kvu, --no-kv-unified` | use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)<br/>(env: LLAMA_ARG_KV_UNIFIED) |
|
||||
| `--cache-idle-slots, --no-cache-idle-slots` | save idle slots to the prompt cache on new task, and clear them when using unified KV (default: enabled, requires cache-ram)<br/>(env: LLAMA_ARG_CACHE_IDLE_SLOTS) |
|
||||
| `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
|
||||
|
||||
@@ -594,7 +594,8 @@ struct server_slot {
|
||||
state = SLOT_STATE_IDLE;
|
||||
|
||||
// do not keep context of the child slots - the parent's context is enough
|
||||
if (task->is_child()) {
|
||||
// nor of the stateless slots - no task ever reuses it
|
||||
if (task->is_child() || task->is_stateless()) {
|
||||
prompt_clear();
|
||||
}
|
||||
|
||||
@@ -1489,6 +1490,13 @@ private:
|
||||
batch.init(ctx_tgt, std::max(n_batch, params_base.n_parallel), n_embd);
|
||||
}
|
||||
|
||||
// decision requests rarely repeat a state, the prompt cache is opt-in for decision models
|
||||
if (decision.type != COMMON_DECISION_TYPE_NONE && !params_base.cache_ram_explicit) {
|
||||
SRV_INF("%s", "decision model: prompt cache disabled by default, use `--cache-ram N` to enable it\n");
|
||||
params_base.cache_ram_mib = 0;
|
||||
params_base.cache_idle_slots = false;
|
||||
}
|
||||
|
||||
if (params_base.cache_ram_mib != 0) {
|
||||
if (params_base.cache_ram_mib < 0) {
|
||||
SRV_TRC("prompt cache is enabled, size limit: %s\n", "no limit");
|
||||
@@ -1782,8 +1790,8 @@ private:
|
||||
if (ret) {
|
||||
update_cache = update_cache && prompt_cache;
|
||||
|
||||
// cache prompts only for completion tasks
|
||||
update_cache = update_cache && task.type == SERVER_TASK_TYPE_COMPLETION;
|
||||
// cache prompts only for the tasks that reuse them
|
||||
update_cache = update_cache && !task.is_stateless();
|
||||
|
||||
if (update_cache) {
|
||||
SRV_TRC("%s", "updating prompt cache\n");
|
||||
@@ -3597,9 +3605,7 @@ private:
|
||||
return;
|
||||
}
|
||||
|
||||
const bool is_stateless_task = slot.task->type == SERVER_TASK_TYPE_EMBEDDING || slot.task->type == SERVER_TASK_TYPE_RERANK;
|
||||
|
||||
if (slot.task->params.cache_prompt && !is_stateless_task) {
|
||||
if (slot.task->params.cache_prompt && !slot.task->is_stateless()) {
|
||||
// reuse any previously computed tokens that are common with the new prompt
|
||||
n_past = slot.prompt.tokens.get_common_prefix(input_tokens);
|
||||
|
||||
|
||||
@@ -300,6 +300,11 @@ struct server_task {
|
||||
bool is_child() const {
|
||||
return id_parent != -1;
|
||||
}
|
||||
|
||||
// embedding and rerank tasks always process their whole prompt, their state is never reused
|
||||
bool is_stateless() const {
|
||||
return type == SERVER_TASK_TYPE_EMBEDDING || type == SERVER_TASK_TYPE_RERANK;
|
||||
}
|
||||
};
|
||||
|
||||
struct result_prompt_progress {
|
||||
|
||||
@@ -41,6 +41,31 @@ def test_rerank():
|
||||
assert least_relevant["index"] == 3
|
||||
|
||||
|
||||
@pytest.mark.parametrize("kv_unified", [False, True])
|
||||
def test_rerank_skips_prompt_cache(kv_unified, tmp_path):
|
||||
global server
|
||||
server.n_slots = 2
|
||||
server.cache_ram = 100
|
||||
server.kv_unified = kv_unified
|
||||
server.debug = True
|
||||
server.log_path = str(tmp_path / "server.log")
|
||||
server.start()
|
||||
|
||||
for i in range(4):
|
||||
res = server.make_request("POST", "/rerank", data={
|
||||
"query": f"query {i}",
|
||||
"documents": TEST_DOCUMENTS[i:],
|
||||
})
|
||||
assert res.status_code == 200
|
||||
|
||||
server.stop()
|
||||
|
||||
with open(server.log_path) as f:
|
||||
log = f.read()
|
||||
assert "__TEST_TAG_CACHE_IDLE_SLOTS_ENABLED__" in log
|
||||
assert "__TEST_TAG_CACHE_IDLE_SLOT__" not in log
|
||||
|
||||
|
||||
def test_rerank_tei_format():
|
||||
global server
|
||||
server.start()
|
||||
|
||||
@@ -158,6 +158,32 @@ def test_systemone_shared_prompt():
|
||||
assert abs(res.body["answers"]["angry"]["noul"] - res_single.body["answers"]["angry"]["noul"]) < 0.01
|
||||
|
||||
|
||||
@pytest.mark.parametrize("cache_ram", [None, 100])
|
||||
def test_systemone_prompt_cache(cache_ram, tmp_path):
|
||||
global server
|
||||
server.n_slots = 1
|
||||
server.cache_ram = cache_ram
|
||||
server.debug = True
|
||||
server.log_path = str(tmp_path / "server.log")
|
||||
server.start()
|
||||
|
||||
# the third request comes back to the state of the first one
|
||||
for state in ["Ticket A: " + TEST_STATE, "Ticket B: " + TEST_STATE * 2, "Ticket A: " + TEST_STATE]:
|
||||
res = server.make_request("POST", "/v1/systemone", data={
|
||||
"state": state,
|
||||
"questions": TEST_QUESTIONS,
|
||||
})
|
||||
assert res.status_code == 200
|
||||
|
||||
server.stop()
|
||||
|
||||
with open(server.log_path) as f:
|
||||
log = f.read()
|
||||
# the prompt cache is opt-in for decision models, and restores their states once enabled
|
||||
assert ("__TEST_TAG_CACHE_IDLE_SLOTS_ENABLED__" in log) == (cache_ram is not None)
|
||||
assert ("found better prompt" in log) == (cache_ram is not None)
|
||||
|
||||
|
||||
def test_systemone_images():
|
||||
global server
|
||||
server = ServerPreset.tinyopenjev()
|
||||
|
||||
Reference in new issue
Block a user