clawperf 0.2.2__tar.gz → 0.2.3__tar.gz

{clawperf-0.2.2 → clawperf-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: clawperf
-Version: 0.2.2
+Version: 0.2.3
 Summary: Performance benchmarking tool for LLM Serving backends with multi-turn long-context workloads
 Author: ClawPerf Contributors
 License-Expression: Apache-2.0
@@ -281,15 +281,14 @@ Each simulated user maintains an independent conversation state with its own gro
 
 Each user's context follows this structure:
 
-```
-[System Prefix] [User Prefix] [History] [Current Input]
-```
+![Context model and compaction](docs/context_model.svg)
 
 When context reaches `--max-context-tokens`, append-mode compaction fires:
 
 1. The base context (system + user prefix + input, without history) is checked first. If it already exceeds the limit, compaction is skipped and the turn is marked as `context_overflow` — this prevents infinite compaction loops.
 2. Otherwise, history is cleared and the user prefix grows by `--compaction-prefix-increment` tokens.
 3. New random content fills the enlarged user prefix.
+4. If the grown base still exceeds the limit, the prefix growth is **reverted** (history cleared only) so the user isn't permanently trapped in overflow.
 
 This simulates how real LLM serving systems handle context overflow with prefix caching.
 
@@ -297,6 +296,8 @@ This simulates how real LLM serving systems handle context overflow with prefix
 
 The mock server simulates vLLM's KV-block prefix cache using a trie:
 
+![Prefix cache reuse across turns](docs/prefix_cache.svg)
+
 - **HBM trie**: Represents GPU KV cache. Queried first for longest prefix match. Always updated after every request (mimicking vLLM storing all KV blocks regardless of hit/miss).
 - **External trie**: Represents CPU/disk prefix cache. Queried on HBM miss. Also always updated after every request.
 - **Token-level hit rate**: `prefix_cache_hit_tokens / prefix_cache_query_tokens` — the fraction of prompt tokens that reuse cached KV blocks. This is the meaningful metric; request-level (binary) hit rate is not reported.
@@ -310,6 +311,8 @@ The mock server simulates vLLM's KV-block prefix cache using a trie:
 
 ## Architecture
 
+![Benchmark pipeline](docs/architecture.svg)
+
 ClawPerf reuses EvalScope's core perf components:
 
 - **AioHttpClient**: Async HTTP with streaming, proper timeout/connector config

{clawperf-0.2.2 → clawperf-0.2.3}/README.md

@@ -246,15 +246,14 @@ Each simulated user maintains an independent conversation state with its own gro
 
 Each user's context follows this structure:
 
-```
-[System Prefix] [User Prefix] [History] [Current Input]
-```
+![Context model and compaction](docs/context_model.svg)
 
 When context reaches `--max-context-tokens`, append-mode compaction fires:
 
 1. The base context (system + user prefix + input, without history) is checked first. If it already exceeds the limit, compaction is skipped and the turn is marked as `context_overflow` — this prevents infinite compaction loops.
 2. Otherwise, history is cleared and the user prefix grows by `--compaction-prefix-increment` tokens.
 3. New random content fills the enlarged user prefix.
+4. If the grown base still exceeds the limit, the prefix growth is **reverted** (history cleared only) so the user isn't permanently trapped in overflow.
 
 This simulates how real LLM serving systems handle context overflow with prefix caching.
 
@@ -262,6 +261,8 @@ This simulates how real LLM serving systems handle context overflow with prefix
 
 The mock server simulates vLLM's KV-block prefix cache using a trie:
 
+![Prefix cache reuse across turns](docs/prefix_cache.svg)
+
 - **HBM trie**: Represents GPU KV cache. Queried first for longest prefix match. Always updated after every request (mimicking vLLM storing all KV blocks regardless of hit/miss).
 - **External trie**: Represents CPU/disk prefix cache. Queried on HBM miss. Also always updated after every request.
 - **Token-level hit rate**: `prefix_cache_hit_tokens / prefix_cache_query_tokens` — the fraction of prompt tokens that reuse cached KV blocks. This is the meaningful metric; request-level (binary) hit rate is not reported.
@@ -275,6 +276,8 @@ The mock server simulates vLLM's KV-block prefix cache using a trie:
 
 ## Architecture
 
+![Benchmark pipeline](docs/architecture.svg)
+
 ClawPerf reuses EvalScope's core perf components:
 
 - **AioHttpClient**: Async HTTP with streaming, proper timeout/connector config

{clawperf-0.2.2 → clawperf-0.2.3}/src/clawperf/__init__.py

@@ -5,4 +5,4 @@ and adds multi-turn long-context workloads with append-mode compaction,
 user arrival scheduling, and system metrics polling.
 """
 
-__version__ = "0.2.2"
+__version__ = "0.2.3"

{clawperf-0.2.2 → clawperf-0.2.3}/src/clawperf.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: clawperf
-Version: 0.2.2
+Version: 0.2.3
 Summary: Performance benchmarking tool for LLM Serving backends with multi-turn long-context workloads
 Author: ClawPerf Contributors
 License-Expression: Apache-2.0
@@ -281,15 +281,14 @@ Each simulated user maintains an independent conversation state with its own gro
 
 Each user's context follows this structure:
 
-```
-[System Prefix] [User Prefix] [History] [Current Input]
-```
+![Context model and compaction](docs/context_model.svg)
 
 When context reaches `--max-context-tokens`, append-mode compaction fires:
 
 1. The base context (system + user prefix + input, without history) is checked first. If it already exceeds the limit, compaction is skipped and the turn is marked as `context_overflow` — this prevents infinite compaction loops.
 2. Otherwise, history is cleared and the user prefix grows by `--compaction-prefix-increment` tokens.
 3. New random content fills the enlarged user prefix.
+4. If the grown base still exceeds the limit, the prefix growth is **reverted** (history cleared only) so the user isn't permanently trapped in overflow.
 
 This simulates how real LLM serving systems handle context overflow with prefix caching.
 
@@ -297,6 +296,8 @@ This simulates how real LLM serving systems handle context overflow with prefix
 
 The mock server simulates vLLM's KV-block prefix cache using a trie:
 
+![Prefix cache reuse across turns](docs/prefix_cache.svg)
+
 - **HBM trie**: Represents GPU KV cache. Queried first for longest prefix match. Always updated after every request (mimicking vLLM storing all KV blocks regardless of hit/miss).
 - **External trie**: Represents CPU/disk prefix cache. Queried on HBM miss. Also always updated after every request.
 - **Token-level hit rate**: `prefix_cache_hit_tokens / prefix_cache_query_tokens` — the fraction of prompt tokens that reuse cached KV blocks. This is the meaningful metric; request-level (binary) hit rate is not reported.
@@ -310,6 +311,8 @@ The mock server simulates vLLM's KV-block prefix cache using a trie:
 
 ## Architecture
 
+![Benchmark pipeline](docs/architecture.svg)
+
 ClawPerf reuses EvalScope's core perf components:
 
 - **AioHttpClient**: Async HTTP with streaming, proper timeout/connector config