Files
llama.cpp/tools/server/server-slot-sched.h
T
LumpiastyandClaude Opus 5.5 7a233df171 server: disk prompt cache, save slots on shutdown and restore after a restart
llama-swap SIGTERMs a model when the owner switches models (only one large
model fits in RAM), and switching back re-prefills the whole conversation:
about 150 s for a 30k-token chat on the 35B. With --cache-disk-path DIR,
each slot with a prompt is saved to DIR on shutdown. After the restart, a
request whose prompt starts with a saved prompt is restored from the file,
so only the suffix is processed. Off by default.

Save, in server_context::start_loop() after the loop returns:
- disk_cache_collect() copies the state of each slot with >= 256 text tokens
  to host memory (llama_state_seq_get_data_ext, flags NONE) while the device
  buffers still exist. A cold server reads its released state through
  llama_context_restore_memory_host(), with no device use and no extra copy.
- vram_share_exit() then frees the VRAM and releases the GPU token as before,
  so only the D2H copy delays the token release, never the disk IO.
- disk_cache_flush() writes the files. A slot is skipped if its KV does not
  cover its token list (seq_pos_max + 1 != n_tokens: the last batch was not
  decoded) or if the same token list is already on disk.

File (server_prompt_disk_cache, server-task.cpp):
- A 48-byte header (magic, version, key hash, sizes, payload hash), then a
  payload of 8-byte padded sections: tokens, checkpoint table, state, the
  checkpoint states. The hash is FNV-1a over the 64-bit words of the payload.
- Plain fwrite to <name>.tmp, fdatasync, POSIX_FADV_DONTNEED, rename. The read
  is a plain fread plus DONTNEED. There is no mmap: upstream #28092 livelocked
  the kernel on AMD + Vulkan by DMA-ing into a MAP_SHARED file.
- The index is built at startup from the headers. A bad magic, a size that
  does not match the header, or an inconsistent table means the file is
  removed. A hash mismatch at load also removes it, and nothing is restored.
  Other keys and other versions are kept. A .tmp older than 1 h is removed.
- --cache-disk-size N MiB (default 4096, 0 = no limit): before a write, the
  least recently used *.lpcd files of every model in DIR are removed until the
  new file fits; a restore updates the file mtime. Entries the new prompt
  extends are dropped after the write and do not count against the limit.
  A write also needs its size + 1 GiB free on the file system. A write error
  or a full disk is logged; the file is not saved and the exit stays clean.

Key: the build commit, the model path + size + mtime, LoRA and control
vectors, n_ctx, n_ctx_seq, n_seq_max, n_ubatch, kv_unified, swa_full, the
K/V types, the flash-attn type, and the RoPE/YaRN parameters.

Restore, in get_available_slot() after the RAM cache:
- Only while warm (holding the GPU token), with cache_prompt on and no media
  tokens. The file must reuse at least 256 more tokens than the slot holds.
- Reusable means one of two things. Either the saved token list is an exact
  prefix of the new prompt, or (for a recurrent/SWA model) a saved checkpoint
  at or below the common prefix can be used. A plain-KV model can be cut back
  to any prefix. Exact prefix alone is not enough for OpenWebUI on Qwen: the
  template rewrites the previous assistant turn (the reasoning is dropped), so
  the saved list is almost never a prefix of the next turn. The checkpoints
  are exact snapshots, and update_slots() rolls back to one exactly as it does
  for the RAM cache. They are restored with id_task -1, so one that fails to
  load means a re-prefill, not an abort.
- The slot's current prompt goes to the RAM cache first. A set_data failure
  clears the slot and removes the file.

Not supported with speculative decoding (warning, disabled): the draft state
is not saved. Only the slots are saved, not the RAM cache entries.

Measured on the pod, CPU only, Qwen3.5-0.8B (hybrid, 6 of 24 layers attention):
- 573-token chat: 64.5 MiB file with 2 checkpoints, copy 18 ms, write 171 ms,
  exit 314 ms (125 ms without the cache). After the restart: read 166 ms,
  restore 3 ms, turn 2 prompt_n 58 of 596 in 462 ms against a 2913 ms full
  prefill. The answer is identical to a server that never restarted.
- 12k tokens: 200.7 MiB, copy 110 ms, write 350 ms, exit 641 ms, peak RSS
  +162 MiB during the exit save. After the restart: read 307 ms, restore 18 ms,
  turn 2 prompt_n 35 of 12188 in 1.2 s against a 72 s prefill.
- The /root PVC writes at about 0.66 GB/s with fdatasync. /tmp (the node
  overlay) stalled once: 5.8 s for 64.5 MiB.
llama-swap v239 waits for the old process to exit before it starts the next
model, and SIGKILLs a process that is still there after 30 s. So the exit
save adds to the model switch time: about 1 s per GB on the PVC.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-21 19:39:56 +01:00

171 lines
5.6 KiB
C++

#pragma once
#include "server-common.h"
#include "server-task.h"
#include <list>
#include <memory>
#include <vector>
// which prompt goes to which slot: slot prompt state, prompt cache and slot selection
// what a cached prompt can give to a new prompt
struct server_prompt_reuse {
int32_t n_past = 0; // tokens to keep
llama_pos pos_next = 0;
// restore this checkpoint first, nullptr when the live state is used as it is
std::shared_ptr<const common_prompt_checkpoint> ckpt;
};
struct server_prompt {
server_tokens tokens;
// a checkpoint does not change after it is created, so the prompt cache shares it with the slot instead of a copy
std::list<std::shared_ptr<const common_prompt_checkpoint>> checkpoints;
// the lora config that the tokens were processed with, it follows the state to the prompt cache and back
// TODO: slot files do not store it
std::vector<common_adapter_lora_info> lora;
void clear() {
tokens.clear();
checkpoints.clear();
}
int n_tokens() const {
return tokens.size();
}
server_prompt clone() const {
return server_prompt {
tokens.clone(),
checkpoints,
lora,
};
}
// n_prefix: tokens shared with the new prompt, must be > 0
// has_new_tokens: the new prompt is longer than n_prefix
// n_swa: 0 if a state must end right before the next token (no SWA, or recurrent)
// pos_min: min position of the live state
// n_past == 0 without a checkpoint means that nothing can be reused
server_prompt_reuse get_reuse(int32_t n_prefix, bool has_new_tokens, int32_t n_swa, llama_pos pos_min) const;
};
struct server_prompt_data {
common_state_data main;
common_state_data drft;
size_t size() const {
return main.size() + drft.size();
}
};
struct server_prompt_cache_state {
server_prompt prompt;
server_prompt_data data;
size_t size() const {
size_t res = data.size();
for (const auto & ckpt : prompt.checkpoints) {
res += ckpt->size();
}
return res;
}
};
struct server_prompt_cache {
server_prompt_cache(int32_t limit_size_mib, size_t limit_tokens) {
this->limit_size = 1024ull*1024ull*(limit_size_mib < 0 ? 0 : limit_size_mib);
this->limit_tokens = limit_tokens;
}
std::list<server_prompt_cache_state> states;
// in bytes, 0 = no limit
size_t limit_size = 0;
// in tokens, 0 = no limit
size_t limit_tokens = 0;
size_t size() const;
size_t n_tokens() const;
server_prompt_cache_state * alloc(const server_prompt & prompt, size_t state_size_main, size_t state_size_drft);
// lora_new: the lora config of the new prompt, the states made with another config are skipped
bool load(server_prompt & prompt, const server_tokens & tokens_new, const std::vector<common_adapter_lora_info> & lora_new, llama_context * ctx_tgt, llama_context * ctx_dft, int32_t id_slot);
void update();
};
// prompt states saved as files, to reuse a prompt after a server restart
struct server_prompt_disk_entry {
std::string path;
llama_tokens tokens;
std::vector<int64_t> ckpt_n_tokens;
};
struct server_prompt_disk_cache {
// key: model file and every parameter that changes the state layout, see the server context
server_prompt_disk_cache(const std::string & dir, int32_t limit_size_mib, const std::string & key);
std::string dir;
uint64_t key_hash;
size_t limit_size;
// entries of this key only, read from the file headers at startup
std::vector<server_prompt_disk_entry> entries;
// tokens of tokens_new that a state of `tokens` with checkpoints at `ckpt_n_tokens` can reuse
// a recurrent or SWA state is reusable only up to an exact prefix or a checkpoint
static size_t n_reuse(const llama_tokens & tokens, const std::vector<int64_t> & ckpt_n_tokens, const llama_tokens & tokens_new, bool can_trim);
// index of the entry that reuses the most of tokens_new, and at least n_min more than n_have; -1 if none
int find(const llama_tokens & tokens_new, size_t n_have, size_t n_min, bool can_trim) const;
bool has(const llama_tokens & tokens);
// read and verify entry i; a bad file is removed
bool load(int i, server_prompt & prompt, common_state_data & data, bool has_mtmd);
// write atomically, after evicting the oldest files in dir over the size limit
bool save(const server_prompt & prompt, const common_state_data & data);
void remove(int i);
};
// what the scheduler needs to know about a slot
struct server_slot_info {
int id = -1;
const server_task * task = nullptr; // only for logging
bool is_processing = false;
int64_t t_last_used = -1;
const server_prompt * prompt = nullptr;
};
struct server_slot_plan {
// -1 if no slot is available; a busy slot if the task asks for it by id, the caller defers the task
int id_slot = -1;
// save the slot state to the prompt cache, then load the best match for the task
bool update_cache = false;
};
// decides only, the caller applies the plan
struct server_slot_sched {
float slot_prompt_similarity = 0.0f;
// idle slots that lose the least by being reused: empty slots first, then the least recently used ones
std::vector<int> get_free_slots(const std::vector<server_slot_info> & slots, size_t n_slots_needed, int exclude_id_slot) const;
// lora: the lora config of the task
server_slot_plan select(const std::vector<server_slot_info> & slots, const server_task & task, const std::vector<common_adapter_lora_info> & lora) const;
};