cfahlgren1/hf-mem-traces

hf-mem traces

Demo copy of alvarobartt/hf-mem-traces. Trace files are reordered as traces/<prompt>/<harness>/ and Codex traces include user_message events, to show all harnesses in the dataset page preview.

Setting	Value
Run	release-a5d5e4c7-20261107
hf-mem revision	a5d5e4c
Skill SHA-256	f5e8976964d372b3c1f1fdf879317e6fdc00d04688d7ef2542ba78dd787bcbe1
Judge / effort	codex/gpt-6.1-terra / high
Harness versions	codex codex-cli 0.167.0; pi 1.5.0

Prompts

ID	Prompt
P1	Can I run Qwen/Qwen3.8-27B on a single RTX 4090 (24 GB) with a 32K context? If not, what's the longest context that would fit, and how many 4090s would I need for 32K?
P2	I have a 16 GB GPU and want to run unsloth/Qwen3.8-27B-GGUF with llama.cpp at an 8K context. Which quantization gives me the best quality that still fits?
P3	I want to serve deepseek-ai/DeepSeek-V4.1-Flash with vLLM on a node with 8x H100 80GB, with a 128K context and 4 concurrent requests. Will it fit, how much memory will each GPU use, and do I need anything else on the machine?
P4	How much VRAM do I need to run google/gemma-4-31B-it?
P5	Will black-forest-labs/FLUX.1-dev fit on my 24 GB GPU?
P6	What's the memory footprint of google/embeddinggemma-2 for inference?
P7	Does meta-llama/Llama-3.1-8B-Instruct fit on an L4 (24 GB) with the full 128K context? Would an FP8 KV cache help?

Summary

Output tokens include reasoning. Cost estimates use input/output rates only, regardless of harness. Only completed runs are published.

Harness / model	Correct	Missed	Errors	Input tokens	Output tokens	Estimated agent USD	Estimated judge USD
codex/gpt-6.1-sol	6	1	0	1,241,831	10,895	2.5926*	0.4254*
codex/gpt-6.1-terra	7	0	0	1,187,076	6,718	1.2207*	0.3202*
codex/gpt-5.6-sol	7	0	0	1,426,559	21,695	6.1401*	0.4692*
pi/gpt-6.1-sol	7	0	0	103,891	6,064	0.2684*	0.1998*
pi/gpt-6.1-terra	7	0	0	68,780	3,173	0.0846*	0.1790*
pi/gpt-6-luna	6	1	0	64,311	5,399	0.0091*	0.1816*
pi/gpt-5.6-sol	6	1	0	119,677	12,902	0.7367*	0.1962*

* OpenAI rates from OpenAI pricing.

** Anthropic rates from Anthropic pricing.

Total estimated cost: $13.0238 ($11.0524 agents, $1.9715 judging).

gpt-6.1-sol with codex

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	89,572	1,007	The answer correctly identifies the original BF16 checkpoint, no fitting context on one 4090, and four GPUs for 32K. Its memory estimates and assumptions match the reference.
P2	❌	299,338	2,764	The main recommendation disables the default projector and relies on unverified manual subtraction and a Q8 cache estimate for UD-Q4_K_S. With the projector retained, the verified choices are UD-Q3_K_XL with FP16 cache or the tight-fitting UD-IQ4_XS with Q8 cache.
P3	✅	409,643	2,968	The answer correctly reports that the requested workload fits at about 55.49 GiB per GPU and identifies the 188.83 GiB of offloaded weights requiring free host RAM.
P4	✅	121,738	1,066	The answer correctly reports the full-context BF16 footprint and labels the shorter-context and quantized alternatives, including their GPU and host RAM requirements.
P5	✅	113,506	834	The answer correctly counts the full 31.42 GiB BF16 pipeline, distinguishes weights from peak inference memory, and offers quantization and CPU offloading as alternatives.
P6	✅	84,861	979	The 1.39 GiB BF16 estimate matches the reference, and the answer clearly distinguishes stored weights from peak inference memory.
P7	✅	123,173	1,277	The answer correctly assesses both caches at 131,072 tokens on one L4. Its memory estimates and context limits match the references, and it identifies quantized weights as an alternative.

gpt-6.1-terra with codex

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	229,645	1,182	The answer correctly distinguishes the original BF16 checkpoint from the quantized alternative, gives the original model's zero-fit context and four-GPU estimate, and labels the verified Q4_K_M estimates and CPU offload.
P2	✅	235,402	1,161	UD-Q3_K_XL fits at 8K with the default F16 cache and projector, while UD-IQ4_XS exceeds the usable budget under those assumptions. The answer correctly notes CPU-offloaded weights and their RAM requirement. The trace also shows successful hf-mem skill discovery and use.
P3	✅	168,494	1,187	The estimate matches the reference: it fits at 55.5 GiB per GPU for four 128K sequences, with 188.8 GiB of pinned host memory required. The answer also identifies the FP8 cache assumption and estimator limitations.
P4	✅	149,195	926	The answer distinguishes the original BF16 model from the quantized alternative, states the context lengths, and matches the estimates. It also identifies CPU-offloaded weights for the 4-bit option.
P5	✅	125,388	646	The answer correctly counts the full pipeline, reports about 31.4 GiB of BF16 weights, and distinguishes that estimate from generation memory. It also identifies quantization and CPU offloading as alternatives for a 24 GB GPU.
P6	✅	198,598	821	The answer correctly reports about 1.39 GiB of BF16 weights and distinguishes that lower bound from peak inference memory, which the reference does not estimate.
P7	✅	80,354	795	Both full-context estimates and the one-L4 fit conclusions match the references. FP8 halves the cache but still does not fit.

gpt-5.6-sol with codex

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	263,157	4,338	The answer correctly distinguishes the requested BF16 checkpoint from quantized alternatives and gives the reference result of no fitting context on one 4090 and four GPUs for 32K.
P2	✅	146,916	4,143	IQ4_XS with Q8_0 KV cache is explicitly verified to fit at 8K with the default projector and a very small margin. The Q3_K_XL alternative and its estimate are also supported.
P3	✅	690,449	7,262	The answer uses a verified 4,096-token scheduler cap and correctly reports fit.required divided by eight as about 53.2 GiB per GPU. It also identifies the 188.8 GiB of pinned host RAM required for offloaded weights.
P4	✅	89,451	1,794	The BF16 estimates and maximum-context breakdown match the references, and the recommendations distinguish shorter contexts from the full window.
P5	✅	82,899	1,203	The answer counts the full pipeline, correctly reports that native BF16 does not fit, and distinguishes weight size from runtime memory.
P6	✅	70,246	1,344	The answer correctly reports 1.39 GiB of BF16 weights and distinguishes that from its approximate inference budget. Actual peak memory depends on workload and runtime; the reference does not model overhead.
P7	✅	83,441	1,611	The answer correctly assesses both cache types at full context and matches the reference estimates, including the maximum contexts and one-GPU fit conclusions.

gpt-6.1-sol with pi

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	12,818	806	The answer correctly identifies that the original BF16 checkpoint does not fit on one 4090 at any context length and estimates four 4090s for 32K. Quantization and CPU offloading are clearly labeled as separate configurations.
P2	✅	36,096	978	The default-cache recommendation matches the reference, and the higher-quality Q8_0 cache alternative is explicitly identified and verified by a tool result. Both estimates include the projector and account for CPU-offloaded weights.
P3	✅	22,197	1,946	The answer correctly reports a fit at 128K context and four requests, calculates 55.5 GiB per GPU from fit.required, and identifies the 188.8 GiB of offloaded weights and free host RAM needed.
P4	✅	12,237	787	The BF16 estimates match the references, and the answer identifies the full 256K default context and labels shorter-context and quantized alternatives.
P5	✅	8,604	507	The answer counts the full pipeline at 31.42 GiB, correctly says BF16 weights do not fit, and distinguishes weight memory from inference overhead.
P6	✅	5,665	328	The answer correctly reports the BF16 weight footprint and makes clear that peak inference memory includes unestimated runtime overhead. The 4 GiB suggestion is labeled as unverified.
P7	✅	6,274	712	The answer correctly assesses both BF16 and FP8 caches at 131,072 tokens, and its memory estimates and context limits match the references.

gpt-6.1-terra with pi

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	11,730	587	The answer correctly identifies that the BF16 checkpoint cannot fit on one 4090 at any context length and estimates four GPUs for 32K. The per-GPU figure and video profiling warning are supported by the tool results.
P2	✅	26,786	650	The Q8_0 cache estimate is verified and retains the default projector. The answer also gives the fitting F16-cache alternative and notes CPU-offloaded weights.
P3	✅	6,283	561	The answer correctly reports the fit, per-GPU requirement, FP8 cache, and pinned host-memory requirement for the requested workload.
P4	✅	6,091	413	The BF16 estimates match the reference, and the answer identifies the full 262,144-token context and its assumptions.
P5	✅	5,941	244	The answer correctly counts the full BF16 pipeline at 31.4 GiB and distinguishes weights from peak inference memory. It gives appropriate offloading and quantization alternatives.
P6	✅	5,673	245	The BF16 weight size and parameter count match the reference, and the answer correctly distinguishes stored weights from peak inference memory.
P7	✅	6,276	473	Both full-context estimates and the conclusion match the independent references. FP8 halves the KV cache but still exceeds one L4.

gpt-6-luna with pi

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	9,164	937	The answer correctly identifies that the BF16 checkpoint cannot fit at any context length on one 4090 and that 32K requires four GPUs. Its memory figures and video warning match the estimate.
P2	✅	18,559	989	Q3_K_XL fits at 8K with the default F16 KV cache and projector; IQ4_XS does not under those assumptions.
P3	❌	9,784	1,588	The fit, per-GPU estimate, and free host RAM requirement are correct. However, the storage estimate omits the 188.8 GiB of offloaded weights; the checkpoint needs about 475 GiB, not 286 GiB.
P4	✅	8,817	461	The estimates match the supplied references, and the answer identifies the context, batch size, and BF16 cache assumptions.
P5	✅	5,980	439	The answer counts the full pipeline, correctly reports about 31.4 GiB of BF16 weights, and distinguishes that estimate from runtime memory.
P6	✅	5,702	368	The answer matches the reference weight estimate and distinguishes it from peak inference memory, for which no overhead was modeled.
P7	✅	6,305	617	Both full-context estimates and fit conclusions match the references, including the benefit and limitation of FP8 KV cache.

gpt-5.6-sol with pi

Prompt	Result	Input tokens	Output tokens	Comment
P1	✅	18,413	1,369	The answer correctly distinguishes the native BF16 checkpoint from quantized alternatives and reports zero fitting context on one GPU and four GPUs for 32K.
P2	❌	32,429	2,324	The quantization estimates and verified Q8 cache alternative are correct, but the answer calls Q3 a full-GPU choice while the estimate offloads about 0.51 GiB of weights to CPU. It should mention the required free host RAM.
P3	✅	29,979	5,149	The answer correctly reports the fit, divides the required memory across eight GPUs, and identifies the need for 188.83 GiB of free pinned host RAM.
P4	✅	13,839	1,386	The native BF16 estimates match the reference, and the full-context FP8 cache alternative is explicitly labeled and supported by a tool result.
P5	✅	8,820	607	The answer correctly counts the full BF16 pipeline at 31.42 GiB, concludes it will not fit, and notes that runtime overhead requires additional memory.
P6	✅	9,861	1,059	The BF16 weight size and parameter count match the reference. The answer distinguishes weight storage from peak inference memory, though its practical VRAM estimate is approximate and not modeled by hf-mem.
P7	✅	6,336	1,008	Both full-context estimates and fit conclusions match the independent references. The answer correctly explains that FP8 halves KV-cache memory but still does not fit on one L4.