mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-07-17 08:28:07 -05:00
68 lines
7.2 KiB
Plaintext
68 lines
7.2 KiB
Plaintext
# solvers.ServingModel { #mlsysim.solvers.ServingModel }
|
|
|
|
```python
|
|
solvers.ServingModel()
|
|
```
|
|
|
|
Analyzes the two-phase LLM serving lifecycle: Pre-fill vs. Decoding.
|
|
|
|
LLM inference is not a single mathematical operation; it is a stateful
|
|
process with two distinct physical regimes (Compute-bound Pre-fill and
|
|
Memory-bound Decoding).
|
|
|
|
Literature Source:
|
|
1. Pope et al. (2023), "Efficiently Scaling Transformer Inference."
|
|
2. Agrawal et al. (2024), "Sarathi-Serve" (chunked prefill scheduling).
|
|
3. Patel et al. (2024), "Splitwise" and Zhong et al. (2024),
|
|
"DistServe" (prefill/decode disaggregation).
|
|
|
|
## Methods
|
|
|
|
| Name | Description |
|
|
| --- | --- |
|
|
| [solve](#mlsysim.solvers.ServingModel.solve) | Solves for LLM serving performance. |
|
|
|
|
### solve { #mlsysim.solvers.ServingModel.solve }
|
|
|
|
```python
|
|
solvers.ServingModel.solve(
|
|
model,
|
|
hardware,
|
|
seq_len,
|
|
batch_size=1,
|
|
precision='fp16',
|
|
efficiency=0.5,
|
|
decode_hardware=None,
|
|
network_bandwidth=Q_('100 GB/s'),
|
|
draft_model=None,
|
|
draft_acceptance_rate=0.7,
|
|
cached_prefix_len=0,
|
|
prefill_chunk_tokens=None,
|
|
)
|
|
```
|
|
|
|
Solves for LLM serving performance.
|
|
|
|
#### Parameters {.doc-section .doc-section-parameters}
|
|
|
|
| Name | Type | Description | Default |
|
|
|-----------------------|---------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------|
|
|
| model | TransformerWorkload | The primary model to be served. | _required_ |
|
|
| hardware | HardwareNode | The hardware node for serving (or pre-fill node in disaggregated serving). | _required_ |
|
|
| seq_len | int | Sequence length (context window). | _required_ |
|
|
| batch_size | int | Batch size. | `1` |
|
|
| precision | str | Numerical precision. | `'fp16'` |
|
|
| efficiency | float | Compute efficiency. | `0.5` |
|
|
| decode_hardware | HardwareNode | If provided, models Disaggregated Serving where 'hardware' does pre-fill and 'decode_hardware' does decoding. KV-cache is transferred over the network. | `None` |
|
|
| network_bandwidth | Quantity | Network bandwidth between pre-fill and decode nodes. | `Q_('100 GB/s')` |
|
|
| draft_model | TransformerWorkload | If provided, models Speculative Decoding using this smaller draft model. | `None` |
|
|
| draft_acceptance_rate | float | Expected acceptance rate (0.0 to 1.0) of draft tokens per step. | `0.7` |
|
|
| cached_prefix_len | int | Number of tokens with pre-computed KV-cache (prompt caching / prefix caching). When > 0, the prefill phase only processes (seq_len - cached_prefix_len) new tokens, reducing TTFT proportionally. The full KV-cache (including cached prefix) still occupies memory. Must be < seq_len. | `0` |
|
|
| prefill_chunk_tokens | int | If provided, split new prefill tokens into chunks of at most this size. This estimates a Sarathi-Serve-style chunked-prefill stall proxy: total TTFT keeps the same compute work plus one dispatch tax per chunk, while decode_stall_bound reports the slowest single chunk that can interfere with ongoing decode iterations. It is not a full scheduler simulation. | `None` |
|
|
|
|
#### Returns {.doc-section .doc-section-returns}
|
|
|
|
| Name | Type | Description |
|
|
|--------|---------------|------------------------------|
|
|
| | ServingResult | Serving performance metrics. |
|