SocialLocalMobile commited on
Commit
4432332
·
verified ·
1 Parent(s): 4ae284e

Clarify AgentX setup and simplify README

Browse files
Files changed (1) hide show
  1. README.md +19 -21
README.md CHANGED
@@ -25,14 +25,13 @@ configs:
25
  **agentx-mini** is a smaller workload for benchmarking local LLM servers, derived from
26
  [semianalysisai/cc-traces-weka-062126-256k](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k).
27
  It contains **200 sessions and 9,439 model calls**, including **1,363 subagent calls**.
28
- The original dataset's filters still apply; the changes below are additional.
29
 
30
  Changes from the original dataset:
31
 
32
  - Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
33
- - Divide active request time by **8** and cap ordinary whole-session idle gaps at **1 second**. Keep original gaps over an hour just above the harness's one-hour stream-separation threshold; the runtime idle cap skips those waits.
34
  - Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
35
- - Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps. The agentx-mini traces are model-independent; actual replay overlap depends on server speed.
36
 
37
  ## Stats
38
 
@@ -46,29 +45,22 @@ total_input_tokens: 126,413,728
46
  total_output_tokens: 962,854
47
  ```
48
 
49
- Main turns are top-level calls, following the source dataset's convention.
50
- Five zero-output calls request one token each during replay, bringing requested
51
- output to **962,859 tokens**. Rendered prompts add tokenizer/chat-template overhead.
52
-
53
- Input totals count repeated prefixes. The ideal estimate is **5,499,568 uncached
54
- input tokens** (**95.65% prefix reuse**), assuming a cold, unlimited cache shared
55
- by each session's agents on one model. Actual prefill depends on the server.
56
 
57
  ## Files and reproduction
58
 
59
  `traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
60
  AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
61
 
62
- To rebuild agentx-mini:
63
 
64
  ```sh
65
  pip install huggingface-hub orjson
66
  python distill.py --output rebuilt
67
  ```
68
 
69
- [distill.py](distill.py) pins the source revision and applies deterministic token
70
- and timeline transforms. It needs no model, tokenizer, or replay plugin.
71
-
72
  ## Run agentx-mini with AgentX Harness
73
 
74
  Requires Python 3.11–3.13. Install in a dedicated virtual environment, following
@@ -80,6 +72,18 @@ source .venv/bin/activate
80
  pip install agentx-harness==1.0.7
81
  ```
82
 
 
 
 
 
 
 
 
 
 
 
 
 
83
  1. **Download agentx-mini** into a local directory.
84
  2. **Start your local LLM server** and load your model, for example using
85
  `llama-server`. It must expose an OpenAI-compatible API. Enable parallel
@@ -92,6 +96,7 @@ pip install agentx-harness==1.0.7
92
  downloaded dataset directory (run `pwd` there to get it).
93
 
94
  ```yaml
 
95
  benchmark:
96
  models: [YOUR_SERVED_MODEL_NAME]
97
  tokenizer:
@@ -123,19 +128,12 @@ Run from that directory:
123
  ```bash
124
  agentx profile \
125
  --config benchmark.yaml \
126
- --random-seed 42 \
127
  --output-artifact-dir ./results/agentx-mini
128
  ```
129
 
130
  **Automatic download:** instead of downloading first, set `hf_weka_dataset` to
131
  `SocialLocalMobile/agentx-mini` and run the same command.
132
 
133
- Keep `timing_mode: agentic_replay` and the runtime idle cap to skip retained long
134
- reference gaps. The harness can occasionally exceed the idle cap around joins.
135
-
136
- Call coverage, prefix reuse, and dependencies were validated across all 200
137
- sessions. Also tested with `agentx-harness==1.0.7` and Qwen3-0.6B on `llama-server`.
138
-
139
  ## License
140
 
141
  [Apache 2.0](LICENSE), matching the source dataset published by SemiAnalysis
 
25
  **agentx-mini** is a smaller workload for benchmarking local LLM servers, derived from
26
  [semianalysisai/cc-traces-weka-062126-256k](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k).
27
  It contains **200 sessions and 9,439 model calls**, including **1,363 subagent calls**.
 
28
 
29
  Changes from the original dataset:
30
 
31
  - Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
32
+ - Divide active request time by **8** and shorten idle gaps. Replay uses a **1-second idle cap**.
33
  - Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
34
+ - Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps. Actual replay overlap depends on server speed.
35
 
36
  ## Stats
37
 
 
45
  total_output_tokens: 962,854
46
  ```
47
 
48
+ Main turns are top-level calls. Input totals include repeated prefixes; actual
49
+ prefill depends on the server's prefix caching. Rendered prompts add
50
+ tokenizer/chat-template overhead.
 
 
 
 
51
 
52
  ## Files and reproduction
53
 
54
  `traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
55
  AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
56
 
57
+ To rebuild agentx-mini using [distill.py](distill.py):
58
 
59
  ```sh
60
  pip install huggingface-hub orjson
61
  python distill.py --output rebuilt
62
  ```
63
 
 
 
 
64
  ## Run agentx-mini with AgentX Harness
65
 
66
  Requires Python 3.11–3.13. Install in a dedicated virtual environment, following
 
72
  pip install agentx-harness==1.0.7
73
  ```
74
 
75
+ **Set the block size for agentx-mini.** AgentX 1.0.7 needs an adjustment for this
76
+ dataset's 8-token blocks. In the activated virtual environment, locate the
77
+ installed plugin configuration:
78
+
79
+ ```bash
80
+ python -c 'from pathlib import Path; import aiperf.plugin; print(Path(aiperf.plugin.__file__).with_name("plugins.yaml"))'
81
+ ```
82
+
83
+ Open that file. Under `public_dataset_loader` → `weka_hf` → `metadata`, change
84
+ `default_block_size: 64` to `default_block_size: 8`. This setting applies to all
85
+ `weka_hf` datasets in this environment; reinstalling the harness can reset it.
86
+
87
  1. **Download agentx-mini** into a local directory.
88
  2. **Start your local LLM server** and load your model, for example using
89
  `llama-server`. It must expose an OpenAI-compatible API. Enable parallel
 
96
  downloaded dataset directory (run `pwd` there to get it).
97
 
98
  ```yaml
99
+ random_seed: 42
100
  benchmark:
101
  models: [YOUR_SERVED_MODEL_NAME]
102
  tokenizer:
 
128
  ```bash
129
  agentx profile \
130
  --config benchmark.yaml \
 
131
  --output-artifact-dir ./results/agentx-mini
132
  ```
133
 
134
  **Automatic download:** instead of downloading first, set `hf_weka_dataset` to
135
  `SocialLocalMobile/agentx-mini` and run the same command.
136
 
 
 
 
 
 
 
137
  ## License
138
 
139
  [Apache 2.0](LICENSE), matching the source dataset published by SemiAnalysis