ridgepoint 0.0.0__tar.gz

ridgepoint-0.0.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Isk4R1oT
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

ridgepoint-0.0.0/PKG-INFO

@@ -0,0 +1,44 @@
+Metadata-Version: 2.4
+Name: ridgepoint
+Version: 0.0.0
+Summary: LLM inference sizing: engine-aware VRAM capacity, KV cache, and roofline intervals. Placeholder — under active development.
+Author: Iskariot
+License: MIT
+Project-URL: Homepage, https://github.com/Isk4R1oT/ridgepoint
+Project-URL: Repository, https://github.com/Isk4R1oT/ridgepoint
+Keywords: llm,inference,vram,kv-cache,roofline,gpu,sizing,vllm
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: license-file
+
+# ridgepoint
+
+**LLM inference sizing that models the engine, not just the weights.**
+
+`ridgepoint` answers the question every VRAM calculator gets wrong: *after your serving
+engine (vLLM, SGLang, …) grabs its memory pool, how many concurrent tokens of KV cache
+actually fit — and where does throughput land on the latency/throughput frontier?*
+
+The name is the **ridge point** of the roofline model: the arithmetic intensity where
+decode stops being memory-bandwidth-bound. It's the single number this tool reasons about.
+
+> 🚧 **Early development.** This `0.0.0` release reserves the name. v0 (Rust core + Python
+> bindings via PyO3/maturin, one code path for both a CLI and `pip install ridgepoint`) is
+> in progress.
+
+## Why another calculator?
+
+Most "will it fit?" tools compute `weights + KV < VRAM`. That is the wrong model for paged
+engines: vLLM pre-reserves `gpu_memory_utilization × VRAM` (default 90%) at startup and
+pages KV into that pool. ridgepoint models what actually decides the answer:
+
+- **Engine-aware capacity** — usable KV pool *after* the pre-grab → max concurrent tokens.
+- **Intervals, not points** — throughput sits on a Pareto frontier set by your latency budget.
+- **MLA vs GQA KV geometry** — DeepSeek-style MLA is ~50–100× smaller than the GQA formula predicts.
+- **Real quant bpw** — Q4_K_M is ~4.8 bits/weight, not the label's 4.5.
+- **OOM honesty** — never silently under-predict memory.
+
+## License
+
+MIT

ridgepoint-0.0.0/README.md

@@ -0,0 +1,30 @@
+# ridgepoint
+
+**LLM inference sizing that models the engine, not just the weights.**
+
+`ridgepoint` answers the question every VRAM calculator gets wrong: *after your serving
+engine (vLLM, SGLang, …) grabs its memory pool, how many concurrent tokens of KV cache
+actually fit — and where does throughput land on the latency/throughput frontier?*
+
+The name is the **ridge point** of the roofline model: the arithmetic intensity where
+decode stops being memory-bandwidth-bound. It's the single number this tool reasons about.
+
+> 🚧 **Early development.** This `0.0.0` release reserves the name. v0 (Rust core + Python
+> bindings via PyO3/maturin, one code path for both a CLI and `pip install ridgepoint`) is
+> in progress.
+
+## Why another calculator?
+
+Most "will it fit?" tools compute `weights + KV < VRAM`. That is the wrong model for paged
+engines: vLLM pre-reserves `gpu_memory_utilization × VRAM` (default 90%) at startup and
+pages KV into that pool. ridgepoint models what actually decides the answer:
+
+- **Engine-aware capacity** — usable KV pool *after* the pre-grab → max concurrent tokens.
+- **Intervals, not points** — throughput sits on a Pareto frontier set by your latency budget.
+- **MLA vs GQA KV geometry** — DeepSeek-style MLA is ~50–100× smaller than the GQA formula predicts.
+- **Real quant bpw** — Q4_K_M is ~4.8 bits/weight, not the label's 4.5.
+- **OOM honesty** — never silently under-predict memory.
+
+## License
+
+MIT

ridgepoint-0.0.0/pyproject.toml

@@ -0,0 +1,22 @@
+# Placeholder build (setuptools) to reserve the name on PyPI.
+# The real v0 will switch this to a maturin backend (Rust core + PyO3 bindings).
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "ridgepoint"
+version = "0.0.0"
+description = "LLM inference sizing: engine-aware VRAM capacity, KV cache, and roofline intervals. Placeholder — under active development."
+readme = "README.md"
+requires-python = ">=3.9"
+license = { text = "MIT" }
+authors = [{ name = "Iskariot" }]
+keywords = ["llm", "inference", "vram", "kv-cache", "roofline", "gpu", "sizing", "vllm"]
+
+[project.urls]
+Homepage = "https://github.com/Isk4R1oT/ridgepoint"
+Repository = "https://github.com/Isk4R1oT/ridgepoint"
+
+[tool.setuptools]
+packages = ["ridgepoint"]

ridgepoint-0.0.0/ridgepoint/__init__.py

@@ -0,0 +1,3 @@
+"""ridgepoint — LLM inference sizing (placeholder, under active development)."""
+
+__version__ = "0.0.0"

ridgepoint-0.0.0/ridgepoint.egg-info/PKG-INFO

@@ -0,0 +1,44 @@
+Metadata-Version: 2.4
+Name: ridgepoint
+Version: 0.0.0
+Summary: LLM inference sizing: engine-aware VRAM capacity, KV cache, and roofline intervals. Placeholder — under active development.
+Author: Iskariot
+License: MIT
+Project-URL: Homepage, https://github.com/Isk4R1oT/ridgepoint
+Project-URL: Repository, https://github.com/Isk4R1oT/ridgepoint
+Keywords: llm,inference,vram,kv-cache,roofline,gpu,sizing,vllm
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: license-file
+
+# ridgepoint
+
+**LLM inference sizing that models the engine, not just the weights.**
+
+`ridgepoint` answers the question every VRAM calculator gets wrong: *after your serving
+engine (vLLM, SGLang, …) grabs its memory pool, how many concurrent tokens of KV cache
+actually fit — and where does throughput land on the latency/throughput frontier?*
+
+The name is the **ridge point** of the roofline model: the arithmetic intensity where
+decode stops being memory-bandwidth-bound. It's the single number this tool reasons about.
+
+> 🚧 **Early development.** This `0.0.0` release reserves the name. v0 (Rust core + Python
+> bindings via PyO3/maturin, one code path for both a CLI and `pip install ridgepoint`) is
+> in progress.
+
+## Why another calculator?
+
+Most "will it fit?" tools compute `weights + KV < VRAM`. That is the wrong model for paged
+engines: vLLM pre-reserves `gpu_memory_utilization × VRAM` (default 90%) at startup and
+pages KV into that pool. ridgepoint models what actually decides the answer:
+
+- **Engine-aware capacity** — usable KV pool *after* the pre-grab → max concurrent tokens.
+- **Intervals, not points** — throughput sits on a Pareto frontier set by your latency budget.
+- **MLA vs GQA KV geometry** — DeepSeek-style MLA is ~50–100× smaller than the GQA formula predicts.
+- **Real quant bpw** — Q4_K_M is ~4.8 bits/weight, not the label's 4.5.
+- **OOM honesty** — never silently under-predict memory.
+
+## License
+
+MIT

ridgepoint-0.0.0/ridgepoint.egg-info/SOURCES.txt

@@ -0,0 +1,8 @@
+LICENSE
+README.md
+pyproject.toml
+ridgepoint/__init__.py
+ridgepoint.egg-info/PKG-INFO
+ridgepoint.egg-info/SOURCES.txt
+ridgepoint.egg-info/dependency_links.txt
+ridgepoint.egg-info/top_level.txt

ridgepoint-0.0.0/ridgepoint.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

ridgepoint-0.0.0/ridgepoint.egg-info/top_level.txt

@@ -0,0 +1 @@
+ridgepoint

ridgepoint-0.0.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+