llm-cachecheck 0.2.1__tar.gz → 0.2.2__tar.gz

{llm_cachecheck-0.2.1 → llm_cachecheck-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: llm-cachecheck
-Version: 0.2.1
+Version: 0.2.2
 Summary: Measure how often the prompt cache changes greedy answers on your LLM server (llama.cpp, plus an experimental two-server mode for vLLM and SGLang)
 Project-URL: Homepage, https://github.com/aditi-p31/cache-divergence-study/tree/main/cachecheck
 Project-URL: Source, https://github.com/aditi-p31/cache-divergence-study
@@ -152,6 +152,13 @@ your results:
 A difference is not a bug in your model. It shows that the answer depends on
 the server's cache state as well as on the request.
 
+## Share your result
+
+At the end of a run the tool prints a short result block. It leaves out your
+server URL and any key. If you would like to share it, open a
+[result issue](https://github.com/aditi-p31/cache-divergence-study/issues/new?template=cachecheck-result.yml)
+and paste the block. Results where no probe differed are just as useful.
+
 ## Background and citation
 
 `cachecheck` adapts the paired cache-on and cache-off design of the study in

{llm_cachecheck-0.2.1 → llm_cachecheck-0.2.2}/README.md

@@ -128,6 +128,13 @@ your results:
 A difference is not a bug in your model. It shows that the answer depends on
 the server's cache state as well as on the request.
 
+## Share your result
+
+At the end of a run the tool prints a short result block. It leaves out your
+server URL and any key. If you would like to share it, open a
+[result issue](https://github.com/aditi-p31/cache-divergence-study/issues/new?template=cachecheck-result.yml)
+and paste the block. Results where no probe differed are just as useful.
+
 ## Background and citation
 
 `cachecheck` adapts the paired cache-on and cache-off design of the study in

{llm_cachecheck-0.2.1 → llm_cachecheck-0.2.2}/cachecheck.py

@@ -41,10 +41,11 @@ import binascii
 import json
 import os
 import sys
+import time
 import urllib.error
 import urllib.request
 
-__version__ = "0.2.1"
+__version__ = "0.2.2"
 
 CITATION = (
     "Patodiya, A. Same Request, Different Answer: Quantization Amplifies "
@@ -54,6 +55,9 @@ CITATION = (
 
 MIN_VALID_SHARE = 0.8
 
+SHARE_URL = ("https://github.com/aditi-p31/cache-divergence-study/issues/new"
+             "?template=cachecheck-result.yml")
+
 # Two-turn probes. Turn 1 asks for a worked answer of a few sentences, so
 # the server writes a useful amount of KV state while generating it.
 # Turn 2 builds on that answer, so it is processed on top of that state.
@@ -305,6 +309,29 @@ def server_props(args):
             "model_file": os.path.basename(props.get("model_path") or "")}
 
 
+def result_block(args, props, summary, when):
+    """A short, copyable summary of the run. It leaves out the server URL
+    and any key, so it is safe to paste in public."""
+    build = props.get("build_info") or "(fill in engine and version)"
+    model = args.model or props.get("model_file") or "(fill in model)"
+    lines = [
+        "---- llm-cachecheck result ----",
+        "tool: llm-cachecheck %s" % __version__,
+        "date (UTC): %s" % when,
+        "mode: %s" % args.mode,
+        "engine build: %s" % build,
+        "model: %s" % model,
+        "GPU or CPU: (fill in)",
+        "probes: %d planned, %d valid, %d differed" % (
+            summary["probes_planned"], summary["probes_valid"],
+            summary["probes_diverged"]),
+        "max tokens: %d, seed: %d, hit check: %s" % (
+            args.max_tokens, args.seed, summary["hit_check"]),
+        "-------------------------------",
+    ]
+    return "\n".join(lines)
+
+
 def status_line(r):
     if not r["valid"]:
         return "inconclusive (%s)" % "; ".join(r["problems"])
@@ -414,6 +441,10 @@ def main(argv=None):
     if args.skip_hit_check:
         print("Cache hits were assumed, not confirmed (--skip-hit-check).")
     print("Report written to %s" % args.out)
+    print()
+    print(result_block(args, props, summary,
+                       time.strftime("%Y-%m-%d %H:%M", time.gmtime())))
+    print("To share this result (optional), open: " + SHARE_URL)
     print("Method: " + CITATION)
 
     if diverged:

{llm_cachecheck-0.2.1 → llm_cachecheck-0.2.2}/test_cachecheck.py

@@ -162,6 +162,16 @@ class CacheCheckTests(unittest.TestCase):
         self.assertIn("DIFFERENT (from character", out)
         self.assertIn("3 of 3 valid probes gave a different", out)
 
+    def test_result_block_is_printed_and_safe_to_share(self):
+        server = self.server(diverge=True)
+        code, report, out = self.run_tool(server)
+        self.assertIn("---- llm-cachecheck result ----", out)
+        self.assertIn("engine build: b-fake", out)
+        self.assertIn("probes: 3 planned, 3 valid, 3 differed", out)
+        self.assertIn(cachecheck.SHARE_URL, out)
+        block = out.split("---- llm-cachecheck result ----")[1]
+        self.assertNotIn(server.url, block)
+
     def test_divergence_in_reasoning_only(self):
         code, report, out = self.run_tool(self.server(diverge_reasoning=True))
         self.assertEqual(code, 1)