Tokenizer Caching¶
SMG implements a two-level tokenizer cache that reduces tokenization overhead for repeated content. It applies where the gateway tokenizes prompts itself: requests to gRPC and ZMQ workers, and /v1/tokenize. HTTP workers tokenize on the engine, so the cache doesn't affect their requests.
Overview¶
L0 Cache (Exact Match)¶
Hash-based O(1) lookup for complete tokenization results. Hits only when an entire input repeats exactly, such as identical prompts in a batch or a resent request.
L1 Cache (Prefix Match)¶
Boundary-aligned prefix matching that tokenizes only the suffix on hit. Ideal for multi-turn conversations with growing context.
Bounded Memory¶
L0 is capped by entry count and by a byte budget; L1 by its own byte budget. An L0 entry holds the whole input and its token IDs, so its size grows with prompt length.
Observable¶
Prometheus counters for lookups (hits and misses), evictions, and reused input bytes, per layer. There are no memory or cache-size metrics.
Why Cache Tokenization?¶
The gateway tokenizes—converts text to token IDs—every request it sends to a gRPC or ZMQ worker. Each tokenization is fast, but the cost grows with prompt length and adds up at scale.
System Prompts¶
Same instructions sent with every request. L1 reuses the tokens up to the last special token the prompts share; L0 helps only when the whole prompt repeats.
Multi-Turn Conversations¶
Growing context with shared prefix. L1 cache tokenizes only new messages.
RAG Applications¶
Queries that share leading context (instructions or documents placed first) reuse it through L1, up to the last special token before the prompts differ.
Batch Processing¶
Identical prompts repeated across a batch hit L0. Prompts that share a template but vary inside it can hit L1 only up to the last special token before the first difference.
Cache Architecture¶
L0 Cache (Exact Match)¶
Gateway-side cache storing complete tokenization results for exact string matches.
- Hash-based O(1) lookup, keyed on the whole input text
- Entry size grows with input length; capped by entry count and a byte budget
- Approximate LRU eviction when full (samples 8 entries and evicts the least recently used)
Best for: Identical requests, repeated batch inputs
L1 Cache (Prefix Match)¶
Gateway-side cache storing tokens at special token boundaries for prefix reuse.
- Tokenize only the suffix on hit
- Cross-request deduplication
- Memory-bounded (configurable)
- Automatic boundary detection
Best for: Multi-turn conversations, growing contexts, incremental content
Each tokenizer SMG loads (from --model-path or --tokenizer-path, or for a gRPC or ZMQ worker's model) gets its own L0 and L1 caches with the configured limits. Tokenizers added through POST /v1/tokenizers are not cached. Decoding is never cached.
Special Token Boundaries (L1)¶
L1 splits inputs right after every special token the tokenizer declares: its BOS, EOS, UNK, SEP, PAD, CLS, and MASK tokens, and every added token marked special. An input without special tokens always misses L1. For example:
| Model Family | Boundary Tokens | Example |
|---|---|---|
| ChatML (Qwen, Yi) | <|im_start|>, <|im_end|> |
Each message boundary |
| Llama 3 | <|begin_of_text|>, <|eot_id|>, <|start_header_id|> |
Text start, turn end |
| GPT | <|endoftext|> |
Document end |
Multi-Turn Conversation Example¶
Consider how caching helps a typical chat application:
Turn 1 (Cold)¶
System: You are a helpful assistant.
User: What is Python?L0: Miss → Full tokenization L1: Miss → Store geometrically spaced checkpoints plus the deepest boundary
Turn 2 (Warm)¶
System: You are a helpful assistant.
User: What is Python?
Assistant: Python is a programming language...
User: How do I install it?L0: Miss (text changed) L1: Hit! → Only tokenize the text after the longest cached boundary; store this turn's deepest boundary and, speculatively, the boundary right after the match
Result: Turn 2 tokenizes only the new messages, not the shared history. The hit also stores a new entry covering Turn 2's prompt up to its own deepest boundary, so Turn 3 matches right after Turn 2's last special token — even when Turn 2's match came from a prefix shared with other conversations, such as a common system prompt.
When more boundaries lie between the match and the deepest one, the hit also stores the boundary right after the match. This is how a shared prefix that spans several special tokens — for example, a long system prompt behind the template's leading token — is learned one boundary per request. The extra entry is speculative: eviction removes it ahead of matched entries unless another request matches it first.
Configuration¶
Model & Tokenizer Paths¶
--model-path¶
HuggingFace model ID or local path to load the tokenizer from.
| Option | --model-path |
|---|---|
| Default | None |
Usage:
# HuggingFace model ID (downloads automatically)
smg --model-path meta-llama/Llama-3.1-8B-Instruct ...
# Local path to model directory
smg --model-path /models/llama-3.1-8b-instruct ...
# Local path to tokenizer.json file
smg --model-path /models/llama-3.1-8b-instruct/tokenizer.json ...When pointing to a local directory, SMG loads the first tokenizer file it finds, in this order:
tekken.json— a Mistral tekken vocabulary. It wins over atokenizer.jsonthe checkpoint ships beside it.tokenizer.json— a HuggingFace tokenizer.- A
vocab.jsonplusmerges.txtpair, whentokenizer_config.jsonnames a Qwen2 tokenizer class. - A tiktoken file (
tiktoken.modelor*.tiktoken). - A
vocab.jsonplusmerges.txtpair without that class.
A directory with none of these is refused with an error naming the accepted files.
A snapshot downloaded from the HuggingFace Hub loads tekken.json first,
then tokenizer.json, then a tiktoken file, then a vocab.json plus
merges.txt pair.
A path to a single .json file loads as a tekken vocabulary when its file
name contains tekken, and as a HuggingFace tokenizer otherwise.
--tokenizer-path¶
Explicit path to a tokenizer file. Overrides --model-path for tokenizer loading.
| Option | --tokenizer-path |
|---|---|
| Default | None |
When to use:
- When the tokenizer is stored separately from the model
- When using a custom tokenizer with a standard model
- When the model directory structure is non-standard
# Load the tokenizer from its own file (--tokenizer-path wins over --model-path)
smg \
--model-path meta-llama/Llama-3.1-8B-Instruct \
--tokenizer-path /custom/tokenizers/llama3-tokenizer.json \
...Chat Templates¶
Chat templates convert structured messages (system, user, assistant roles) into the prompt format expected by specific models. SMG uses Jinja2 templates, the same format used by HuggingFace Transformers.
--chat-template¶
Path to a Jinja2 chat template file.
| Option | --chat-template |
|---|---|
| Default | Auto-discovered from model |
Template discovery priority:
- An explicit template: a model card's
chat_templatepath first, then--chat-template chat_template.jinjain model directorychat_templatefield intokenizer_config.json— a string, or a list of named templates (thedefaultentry renders plain requests, thetool_useentry renders requests that declare tools)- Legacy
chat_template.jsonin model directory (used only whentokenizer_config.jsoncarries no template) - Any other
.jinjafile in model directory
A checkpoint may ship a processor chat_template.json beside the tokenizer's template; requests without tools can render through it. See Chat Templates and Native Renderers.
Template Variables¶
Chat templates use Jinja2 syntax with access to:
| Variable | Description |
|---|---|
messages |
Array of message objects with role and content |
add_generation_prompt |
Boolean to add assistant prompt prefix |
tools |
Optional array of tool definitions |
documents |
Optional array of document context |
bos_token, eos_token, unk_token, pad_token |
The tokenizer's special tokens, when it defines them |
strftime_now |
Function writing the current date and time with a strftime format (see below) |
dict |
Function building a map from a mapping or from key/value pairs, as Python's dict does (see below) |
Keys in a request's chat_template_kwargs are passed to the template as extra variables.
strftime_now matches the function HuggingFace Transformers provides to chat templates: {{ strftime_now('%Y-%m-%d') }} writes today's date in the server's local time, like Python's datetime.now().strftime(format). Every call in one render writes the same instant. An invalid strftime format fails the render with an error naming strftime_now.
To make a prompt that writes the date reproducible, set SOURCE_DATE_EPOCH in SMG's environment to the desired instant as seconds since the Unix epoch (the reproducible-builds convention). SMG reads it once, at the first render, and every render then writes that instant instead of the clock. A value that is not a whole number of seconds is ignored, and the clock is used.
dict matches Python's built-in: dict() builds an empty map, dict(mapping) copies a mapping, and dict(pairs) builds a map from an iterable of key/value pairs — the form templates use to merge maps, as in dict((a | items | list) + (b | items | list)). Keyword arguments are written over the result, and a repeated key keeps its first position with its last value. An argument that is not a mapping or iterable, or an element that is not a two-item key/value pair, fails the render as it would in Python.
Template Examples¶
ChatML (Qwen, Yi)
{%- for message in messages %}
<|im_start|>{{ message.role }}
{{ message.content }}<|im_end|>
{% endfor %}
{%- if add_generation_prompt %}
<|im_start|>assistant
{% endif %}Llama 3
<|begin_of_text|>{% for message in messages %}
<|start_header_id|>{{ message.role }}<|end_header_id|>
{{ message.content }}<|eot_id|>
{% endfor %}
{% if add_generation_prompt %}<|start_header_id|>assistant<|end_header_id|>
{% endif %}L0 Cache Configuration¶
The L0 cache stores complete tokenization results for exact string matches.
--tokenizer-cache-enable-l0¶
Enable the L0 exact match cache.
| Option | --tokenizer-cache-enable-l0 |
|---|---|
| Default | false |
--tokenizer-cache-l0-max-entries¶
Maximum number of entries in the L0 cache.
| Option | --tokenizer-cache-l0-max-entries |
|---|---|
| Default | 10000 |
--tokenizer-cache-l0-max-memory¶
Byte budget for the L0 cache: the texts, token IDs, and per-entry overhead it holds.
| Option | --tokenizer-cache-l0-max-memory |
|---|---|
| Default | 268435456 (256 MB) |
L1 Cache Configuration¶
The L1 cache stores tokenization results at special token boundaries.
--tokenizer-cache-enable-l1¶
Enable the L1 prefix matching cache.
| Option | --tokenizer-cache-enable-l1 |
|---|---|
| Default | false |
--tokenizer-cache-l1-max-memory¶
Maximum memory for the L1 cache in bytes.
| Option | --tokenizer-cache-l1-max-memory |
|---|---|
| Default | 52428800 (50 MB) |
Memory Planning¶
L0 Cache Sizing¶
L0 is capped two ways: by --tokenizer-cache-l0-max-entries and by the --tokenizer-cache-l0-max-memory byte budget. SMG charges each entry its input text plus 4 bytes per token and a small fixed overhead, so entries grow with prompt length: small for a short prompt, multiple megabytes for a large one. An entry that alone would take more than a quarter of the byte budget is not cached at all. When an insert would exceed either bound, SMG evicts approximately least recently used entries (sampling 8 at a time) to make room.
Size --tokenizer-cache-l0-max-entries from the number of distinct whole prompts that actually repeat in your traffic; the byte budget then bounds the memory those entries can take. As with L1, the budget is an estimate of cache contents, not a cap on process memory.
L1 Cache Sizing¶
L1 cache is bounded by total memory:
| Memory | Recommended For |
|---|---|
| 25 MB | Memory-constrained environments |
| 50 MB | Standard deployments (default) |
| 100 MB | Multi-turn conversation heavy |
| 200 MB | Long context applications |
On a miss, L1 stores entries for a selection of the input's special-token boundaries, not all of them: geometrically spaced checkpoints — each stored prefix at least doubling the token length of the one before — plus the input's deepest boundary. Each entry holds the token IDs of the whole prefix up to its boundary. On a hit, L1 stores the prefix up to the input's own deepest boundary when that boundary lies past the match, so steady hit traffic also grows the cache. When further boundaries lie between the match and the deepest, the hit also stores the prefix up to the boundary right after the match, as a speculative entry that eviction removes ahead of matched entries unless another request matches it first.
SMG charges each entry the prefix's length in bytes plus 4 bytes per token, so a miss charges at most a few times the input's own length no matter how many boundaries it has. When new entries would exceed the budget, SMG evicts approximately least recently used entries (sampling 32 at a time) to make room; an entry larger than the whole budget is not stored at all. The budget is an estimate of cache contents, not a cap on process memory.
Recommended Configurations¶
High-Throughput Chat¶
For workloads that resend identical prompts, such as batch jobs and client retries.
smg \
--model-path meta-llama/Llama-3.1-8B-Instruct \
--tokenizer-cache-enable-l0 \
--tokenizer-cache-l0-max-entries 50000Expected: hits on exact repeats of a whole prompt
Multi-Turn Conversations¶
For chat applications with varying conversation lengths.
smg \
--model-path Qwen/Qwen2.5-7B-Instruct \
--tokenizer-cache-enable-l0 \
--tokenizer-cache-l0-max-entries 20000 \
--tokenizer-cache-enable-l1 \
--tokenizer-cache-l1-max-memory 104857600Expected: L0 catches exact repeats, L1 accelerates prefix sharing
Memory-Constrained¶
For deployments with limited memory.
smg \
--model-path meta-llama/Llama-3.1-8B-Instruct \
--tokenizer-cache-enable-l0 \
--tokenizer-cache-l0-max-entries 5000 \
--tokenizer-cache-l0-max-memory 67108864Expected: the entry cap and a 64 MB byte budget bound L0 memory; hits only on exact repeats
✗ No Caching¶
For stateless deployments or when memory is critical.
smg \
--model-path meta-llama/Llama-3.1-8B-Instruct
# Caching is disabled by defaultUse when: Diverse, unique requests dominate
Complete Example¶
Production configuration with tokenizer and caching:
smg \
--worker-urls grpc://worker1:50051 grpc://worker2:50051 \
--policy cache_aware \
--model-path meta-llama/Llama-3.1-70B-Instruct \
--chat-template /templates/llama3.jinja \
--tokenizer-cache-enable-l0 \
--tokenizer-cache-l0-max-entries 25000 \
--tokenizer-cache-enable-l1 \
--tokenizer-cache-l1-max-memory 104857600 \
--host 0.0.0.0 \
--port 8080Monitoring & Observability¶
The gateway exports cache activity on its Prometheus /metrics endpoint:
| Metric | Labels | Meaning |
|---|---|---|
smg_tokenizer_cache_lookups_total |
layer (l0, l1), result (hit, miss) |
One outcome per lookup |
smg_tokenizer_cache_evictions_total |
layer |
Entries removed to make room (clears and replacements excluded) |
smg_tokenizer_cache_reused_bytes_total |
layer |
Input bytes served by hits: whole inputs for L0, matched prefixes for L1 |
The counters are totals across all tokenizers in the process, with no model label, and a disabled layer reports zeros. An L0 hit never reaches L1, and L1 counts inputs without special-token boundaries as misses, so compute each layer's hit ratio against its own lookups:
sum by (layer) (rate(smg_tokenizer_cache_lookups_total{result="hit"}[5m]))
/
sum by (layer) (rate(smg_tokenizer_cache_lookups_total[5m]))See the Metrics Reference for details.
Sizing Signals to Watch¶
Use these signals when tuning --tokenizer-cache-l0-max-entries and
--tokenizer-cache-l1-max-memory:
- A low L0 hit ratio with a steady L0 eviction rate
(
rate(smg_tokenizer_cache_evictions_total{layer="l0"}[5m])) means more distinct inputs arrive than L0 holds; raisemax-entries— and--tokenizer-cache-l0-max-memorywhen large prompts fill the byte budget first — if the workload repeats inputs at all. - L1 evictions mean the prefix cache is at its memory bound. Multi-turn chat traffic with growing context benefits from a larger L1 budget; see L1 Cache Sizing.
smg_tokenizer_cache_reused_bytes_totalseparates many small prefix hits from reuse of large prompts: a high hit ratio with little reused volume saves little tokenization work.- Each L0 entry keeps the full input text and its token IDs, so entry size
grows with prompt length — multi-megabyte entries are possible for large
inputs.
--tokenizer-cache-l0-max-memorybounds their total; watchsmg_allocator_allocated_byteswhen raising it ormax-entries.
Integration with Other Caching Layers¶
Tokenizer caching is part of SMG's three-level caching strategy:
| Layer | What's Cached | Benefit |
|---|---|---|
| Tokenizer L0/L1 | Token IDs | Skip tokenization |
| Router radix tree | Prefix → worker mapping | Consistent routing decisions |
| Worker KV cache | Attention states | Skip prefill computation |
What's Next?¶
Cache-Aware Routing¶
Maximize KV cache hits with prefix-based worker affinity.