40 tok/s at 64K context — 262K context validated — a build report and a practitioner read

ver07 — 09/May/2026 — Des Donnelly — https://www.dd.ie/

Researched, built, written, and validated entirely on the hardware described. The first draft was generated by Gemma 4 26B A4B running locally; subsequent revisions worked from the configuration data captured in the live server logs. The mistakes corrected along the way are noted in the document. Hardware bought at consumer prices, work done from Co Tyrone in North Ireland, with no sponsorship and no early access. Good luck building and kudos to AMD!

1. Introduction

Google announced last week that Gemma 4 has been downloaded over 60 million times since release. A small fraction of those downloads have been turned into running systems. A smaller fraction again on AMD hardware. A smaller fraction still on a 16 GB consumer card. This document is for that small fraction — and for everyone else who is wondering whether it is possible.

It is. The numbers are real, the build is reproducible, and what follows is the working configuration, the traps that cost me time, and an honest assessment of the limits of the resulting system.

A note on perspective: I have been a paying user of frontier cloud LLMs since 2024 and use them daily. That calibration shapes how I read what local hardware delivers, and I will name where that lens matters. The aim is a piece useful to AMD builders, with claims that match what the hardware actually does.

2. The Headline Numbers

Three configurations were tested and validated on this hardware. All three load successfully. Each has its own trade-off and use case.

The 262K figure is the model’s native training context (n_ctx_train = 262144), validated on 16 GB AMD consumer hardware.

Verified environment:

64K @ 40 tok/s for daily use. 262K @ 26.74 tok/s for full native context. The number that matters depends on the task; both are reproducible from the launch commands documented below.

3. The Components That Affect Performance

Only the parts that materially affect Gemma’s speed and behaviour:

The iGPU Detail

On AM5 hardware, ROCm enumerates two devices: the discrete RX 9060 XT (Device 0, gfx1200) and the integrated GPU on the Ryzen die (Device 1, gfx1036). Without an explicit HIP_VISIBLE_DEVICES=0 pin, llama.cpp can spread layers across both, dragging decode speed from ~40 tok/s to single digits. This is the most important single environment variable in the whole build.

4. The Model: Gemma 4 26B A4B

Architectural facts confirmed from the GGUF metadata:

Why This Architecture Fits 16 GB

Three architectural facts work together. Any one alone would not be enough; together they make the build viable:

Mixture of Experts (MoE) reduces active compute. Of 128 experts in the model, 8 fire per token. Of 25.2B total parameters, ~3.8B are active at any moment. The GPU moves roughly 15% of the model’s weight per inference step. This places the practical reasoning capacity in the range of a well-trained 4–8B dense model, rather than a true 26B dense one.

Sliding Window Attention (SWA) caps KV cache growth. In a standard transformer, KV cache scales linearly with context length. With SWA, alternating layers attend to a fixed window of 1,024 tokens regardless of total context. The model metadata confirms an alternating pattern across the 30 layers. The practical result, measured on this build:

Each doubling of context costs roughly 1.6× the KV cache, not 2×. A standard transformer would have shown 2× scaling at every step. This is the architecture’s contribution made visible.

IQ3_M (IQ3_S mix) hits a usable size point. The bartowski quantisation lands at 12.35 GiB at 4.21 bits-per-weight. The naming is slightly misleading — the file is labelled IQ3_M, but llama.cpp reports the actual format as “IQ3_S mix”, with a blend of q5_0, q5_1, q8_0, q5_K, q6_K, iq4_nl, and iq3_s tensors. The mix preserves higher precision where it matters most while compressing aggressively elsewhere.

5. The Working Configuration (64K Production)

The launch command that produced the 40 tok/s headline figure:

HIP_VISIBLE_DEVICES=0 ~/llama.cpp/build-hip/bin/llama-server \
  -m ~/models/gemma-4-26b-a4b-it/*IQ3_M*.gguf \
  --host 127.0.0.1 --port 8090 \
  --n-gpu-layers 99 \
  --ctx-size 65536 \
  --flash-attn on \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --n-cpu-moe 4 \
  --jinja

Flag Breakdown

What Is Not in the Command

No GRUB kernel parameters. Some guides recommend amdgpu.gttsize=16384 or similar for VRAM ‘spillover’ into system memory. This build runs at default kernel parameters — GRUB_CMDLINE_LINUX="" — and never spills. The architectural choices above keep the working set inside VRAM.

No batch-size override. Default n_batch=2048 and n_ubatch=512 are appropriate for this hardware. Earlier guidance that recommended --batch-size 256 came from a pre-flash-attn era and is no longer needed.

Memory Breakdown at 64K

From the live server log at startup:

Model weights on GPU:    11,092 MiB
KV cache (Q8_0):          1,158 MiB  (680 non-SWA + 478 SWA)
Compute buffers:            532 MiB
-------------------------  ------
Total GPU memory:        12,782 MiB
Free GPU memory:          2,801 MiB
CPU offload (4 experts):  2,284 MiB

6. Two Larger Configurations

The same launch command, with selected flags adjusted, gives two further valid configurations.

131K (Long-Context Daily Use)

Change --ctx-size 65536 to --ctx-size 131072. Everything else stays the same.

HIP_VISIBLE_DEVICES=0 ~/llama.cpp/build-hip/bin/llama-server \
  -m ~/models/gemma-4-26b-a4b-it/*IQ3_M*.gguf \
  --host 127.0.0.1 --port 8090 \
  --n-gpu-layers 99 \
  --ctx-size 131072 \
  --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --n-cpu-moe 4 \
  --jinja

Memory at 131K: 13,462 MiB GPU (model 11,092 + KV 1,838 + compute 532), 2,121 MiB free, 2,412 MiB host offload. Decode speed ~37 tok/s, sampled.

262K (Full Native Context)

Two changes from the 131K command: --ctx-size 262144 and --parallel 1. Also --n-cpu-moe 8 to free additional GPU memory for the larger compute buffers and KV cache that 262K requires.

HIP_VISIBLE_DEVICES=0 ~/llama.cpp/build-hip/bin/llama-server \
  -m ~/models/gemma-4-26b-a4b-it/*IQ3_M*.gguf \
  --host 127.0.0.1 --port 8090 \
  --n-gpu-layers 99 \
  --ctx-size 262144 \
  --parallel 1 \
  --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --n-cpu-moe 8 \
  --jinja

Memory at 262K: 13,443 MiB GPU (model 9,579 + KV 2,879 + compute 984), 2,126 MiB free, 4,175 MiB host offload. Decode speed 26.74 tok/s, measured over a 1,712-token output.

The trade-off: --parallel 1 means a single conversational slot. Without it, llama.cpp silently rewrites the requested 262K context down to 512 tokens (see Hard-Won Lessons). For personal use — one user, one conversation at a time — single-slot is the natural configuration anyway. For multi-user serving, the practical ceiling on this hardware is 131K.

A second observation worth recording: compute buffers are not flat across context sizes. At 64K and 131K they hold at 532 MiB; at 262K they grow to 984 MiB. Flash Attention’s working space scales gently with context, just not at the rate the KV cache does.

7. ROCm Setup

A short version of what worked, with the points that caused friction:

# 1. Download the AMD installer (pin exact version)
wget https://repo.radeon.com/amdgpu-install/7.2.1/ubuntu/noble/amdgpu-install_7.2.1.70201-1_all.deb

# 2. Install the installer package
sudo apt install -y ./amdgpu-install_7.2.1.70201-1_all.deb

# 3. Install ROCm — graphics + rocm, no DKMS, no 32-bit
sudo amdgpu-install -y --usecase=graphics,rocm --no-dkms --no-32

# 4. Add yourself to render and video groups
sudo usermod -a -G render,video $USER

# 5. Reboot, then verify
rocminfo | grep -E "Marketing Name|gfx1200"

Note: ROCm 7.2.2 attempted first — the graphics/7.2.2 sub-repository returned HTTP 404 during apt update. Downgrade to 7.2.1 (AMD’s officially supported release for Ubuntu 24.04 Noble) resolved this. Worth verifying repo URLs against repo.radeon.com before following any guide.

Note: On kernel 6.17, the RDNA 4 drivers are in-tree. Installing rocm-dkms creates parallel kernel modules that can conflict on a kernel update. Use --no-dkms.

Note: dmesg requires sudo on Ubuntu 24.04 (kernel.dmesg_restrict=1 is set by default). Older guides that run dmesg without sudo will silently return nothing.

What rocminfo Should Show

After install, both GPUs should be enumerated:

Marketing Name:          AMD Radeon RX 9060 XT
  Name:                    gfx1200
Marketing Name:          AMD Radeon Graphics       (the iGPU)
  Name:                    gfx1036

Both visible is correct. The HIP_VISIBLE_DEVICES=0 pin in your launch command keeps the iGPU out of the inference path.

8. Building llama.cpp

Build the HIP backend from source, targeting gfx1200. The repo has moved — use the new URL:

# Clone (note: ggml-org, not ggerganov)
git clone https://github.com/ggml-org/llama.cpp ~/llama.cpp
cd ~/llama.cpp && git rev-parse HEAD

# Confirm the cmake flag name in your checkout
grep -E "GGML_HIP\b|GGML_HIPBLAS\b" ~/llama.cpp/CMakeLists.txt

# Build
export PATH=/opt/rocm/bin:$PATH
export CMAKE_PREFIX_PATH=/opt/rocm
cmake -S ~/llama.cpp -B ~/llama.cpp/build-hip \
  -DGGML_HIP=ON \
  -DAMDGPU_TARGETS=gfx1200 \
  -DGGML_HIP_NO_VMM=1 \
  -DCMAKE_BUILD_TYPE=Release
cmake --build ~/llama.cpp/build-hip --config Release -j$(nproc)

Note: Current llama.cpp uses -DGGML_HIP=ON. Older guides use -DGGML_HIPBLAS=ON. The grep step above tells you which applies to your checkout. The wrong flag produces a CPU-only binary that reports no HIP devices.

Vulkan Fallback Build

Worth building after the HIP build succeeds. Provides a working alternative if HIP fails on first run.

# Install full SPIR-V toolchain — spirv-tools alone is insufficient
sudo apt -y install libvulkan-dev glslc spirv-tools spirv-headers vulkan-tools

cmake -S ~/llama.cpp -B ~/llama.cpp/build-vulkan \
  -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build ~/llama.cpp/build-vulkan --config Release -j$(nproc)

Note: The Vulkan build will fail at ~47% with: ‘spv’ is not a class, namespace, or enumeration — if spirv-headers is not installed. The spirv-tools package does not pull it in as a dependency.

Downloading the Model

Accept the Gemma 4 licence at huggingface.co/google/gemma-4-26b-a4b-it (Google form, ~2 minutes), generate a HF token, then:

# Set up uv-managed Python venv
uv venv --python 3.12 ~/venvs/ai && source ~/venvs/ai/bin/activate
uv pip install -U "huggingface_hub" hf_xet

# Login (use --token flag; masked-input paste drops tokens silently)
hf auth login --token "hf_PASTE_YOUR_TOKEN_HERE"

# Download IQ3_M quant (~12.35 GiB)
hf download bartowski/google_gemma-4-26B-A4B-it-GGUF \
  --include "*IQ3_M*.gguf" \
  --local-dir ~/models/gemma-4-26b-a4b-it

Note: huggingface-cli is deprecated in HF Hub 1.x. Use hf instead.

9. Running the Server

Wrap the launch in a tmux session so it survives terminal close:

sudo apt -y install tmux  # not pre-installed on minimal Ubuntu 24.04

tmux new-session -d -s llama "HIP_VISIBLE_DEVICES=0 \
  ~/llama.cpp/build-hip/bin/llama-server \
  -m ~/models/gemma-4-26b-a4b-it/*IQ3_M*.gguf \
  --host 127.0.0.1 --port 8090 \
  --n-gpu-layers 99 \
  --ctx-size 65536 \
  --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --n-cpu-moe 4 --jinja \
  2>&1 | tee ~/logs/llama-server-\$(date +%F-%H%M).log"

Verify the Server is Ready

until curl -s http://127.0.0.1:8090/v1/models | grep -q '"id"'; do
  echo "waiting..."; sleep 2
done && echo "READY"

OpenAI-compatible API at http://127.0.0.1:8090/v1, plus a built-in web chat UI at http://127.0.0.1:8090. Both work out of the box.

10. Model Behaviour

Reasoning mode is on by default. Gemma 4 outputs a <think> block before its final answer. The reasoning_content field contains the internal scratchpad; content contains the user-facing response. Prefix prompts with /no_think for direct, fast answers.

finish_reason: “length” instead of “stop”. Reasoning mode consumes tokens before the final answer. Use max_tokens ≥ 500 for short answers, ≥ 2000 for code generation. At max_tokens: 40, the model exhausts its budget on internal reasoning and returns empty content.

Conversation truncated mid-response. The active context limit has been reached. Start a new conversation client-side; no server restart needed.

Training cutoff: approximately January 2025. Release: 02/04/2026.

11. Connecting an IDE: Status as of 09/05/2026

This section is unfinished, deliberately. The state of investigation:

Cursor

Cursor 3.2.16 has a current limitation in which custom local model names do not survive the request pipeline. The Override Base URL setting appears to work in the UI but the model identifier does not route to a locally-served endpoint. Cursor’s primary focus is cloud inference, so a fix may not be near-term.

Antigravity

Google Antigravity allows custom model configuration via .antigravity/config.json, pointing at a local OpenAI-compatible endpoint. The configuration is straightforward (Masaki Hirokawa’s antigravitylab.net articles cover it well). The constraint is performance.

Antigravity is built around an Architect/Builder agent pattern. Every request triggers multiple sequential LLM calls — planning, decomposition, implementation, verification. That works comfortably at cloud speeds (200+ tok/s). At 40 tok/s with prefill on each step, a single ‘explain this function’ turn takes 30–60 seconds. The slowness reflects the architectural fit between the agent paradigm and single-card local hardware, not a configuration error.

VS Code + Continue.dev

Continue.dev is open-source, MIT-licensed, and architected for single-call chat against a custom endpoint. No agent orchestration overhead, designed local-first. It fits the performance profile of this hardware.

I have not yet tested it on this build. The next attempt to close the IDE question will start there. If it works, the integration is solved in 20 minutes. If it does not, that is a useful data point too.

Honest status: the model is running, the API is live, the IDE front-end remains an open problem. Builders who need a working IDE today should plan to test their preferred client end-to-end against their own hardware before relying on it.

12. Hard-Won Lessons

Things worth knowing before you start, drawn from where time was lost:

13. Practitioner Caveats

A section on what local LLM at this hardware tier delivers in practice, written by someone who pays for frontier cloud daily and uses both.

What This System Does Well

Gemma 4 26B A4B at IQ3_M, with 8 active experts of 128 totalling ~3.8B active parameters, is competent in its weight class. It writes a function. It summarises a paragraph. It handles a well-defined classification task. It runs offline, with no API costs, no rate limits, and no data leaving the hardware. For privacy-bound work, narrow tasks, and unmetered iteration it is genuinely useful. None of that is small.

Where Its Limits Show

The active-parameter capacity is in the range of a small dense model, and the effective reasoning depth follows from that. Sustained coherence over long outputs, holding three or four constraints in working memory simultaneously, the kind of inference that frontier cloud models do without thinking — these all show their limits within ten minutes of real use. The Hirokawa article on antigravitylab.net flags this directly: Gemma needs prompts that are dramatically more explicit than what works with frontier models. Users coming from Claude or GPT will need to recalibrate their prompting style.

Three Common Claims, with Context

“Privacy.” Genuinely material for regulated work, confidential client material, or anything that should not pass through a third-party API. Less material for the chatting-about-code, drafting-emails workload that makes up most cloud LLM use under enterprise terms. Worth being clear which category your use sits in.

“No fees.” Operationally true once built; the upfront hardware cost (£1,500+), electricity under load, build and maintenance time, and the opportunity cost of using a less capable tool are all real. £20/month for Claude Pro is in the same financial bracket as a year of electricity for the GPU.

“Frontier-class at home.” Benchmark scores for open models are often within a few points of frontier closed models on standard tests. Real-world capability gaps are wider than the benchmarks suggest. The gap is most visible in sustained reasoning work, less visible in short-form generation.

You can run a 26B MoE at 40 tok/s on a 16 GB consumer card. You can also reach the model’s full native 262K context on the same hardware. Both are hard-won and worth documenting. The model is competent within its weight class. It is not a substitute for frontier cloud reasoning, and that distinction is worth holding in mind when reading any local-LLM literature, including this one.

Where Local LLM Fits

In a portfolio that includes a frontier cloud subscription, local LLM keeps a real chair. It handles work that should not go through a cloud API. It is always available, regardless of internet, rate limits, or service status. It supports unmetered experimentation — trying twenty prompt variations to refine a workflow, learning the dynamics of an open model, building skills against an endpoint with predictable latency. These are real capabilities that did not exist on hardware this affordable two years ago.

Used for what it is rather than what the broader literature sometimes implies, this build is genuinely useful. The point of writing it up is to give other AMD builders the same baseline.

14. Summary

For AMD builders: Pin your devices. Quantise your KV cache to Q8_0 with --flash-attn on. Build from source for gfx1200. Use --no-dkms. Install spirv-headers before the Vulkan build. Use --parallel 1 if you want full 262K context. And if you ever expand to a second NVMe, physically disconnect the first drive before running the installer.

The hardware is ready. The software is catching up. What works, works — and the parts that are still in motion are honestly named here so other builders can plan around them.

15. The Twilight Zone

Two recent announcements from Google point at a near-future where the numbers in this article look conservative. Worth flagging both, while being clear that nothing in this section has been measured on this hardware — only mathematically projected from what others have published.

Multi-Token Prediction Drafters (Released 05/05/2026)

Google’s MTP drafters for the Gemma 4 family use speculative decoding with KV-cache sharing between target and drafter models. The announcement headline cites up to 3× speedup; the actual benchmark chart in the same blog post tells a more nuanced story. The published table covers seven hardware/model combinations:

Mean: 2.41×. Median: 2.5×. Range: 1.5–3.1×. Coefficient of variation: ~24%. That spread is well above measurement noise and tells a real story.

Why the variance is wide: speculative decoding only helps when there is idle compute capacity to spare during single-token decode. The standard inference loop on most hardware is memory-bandwidth-bound rather than compute-bound — the GPU spends most of each token’s time fetching weights from VRAM, while the compute units sit partly idle. MTP captures that idle compute by speculating ahead and verifying in parallel. The speedup is therefore proportional to how much idle compute headroom exists during single-token decode.

That depends on three things:

• Memory bandwidth relative to compute throughput. A100 has ~2 TB/s and ~78 TFLOPS FP16. Mobile hardware is an order of magnitude lower on both. Bandwidth-to-compute ratios vary substantially across this hardware list.
• Model size relative to that ratio. Bigger models create more memory pressure per token. The doubling from 1.5× (26B on A100) to 3.0× (31B on A100) on the same hardware shows the larger model exposing more idle compute that MTP can use.
• Architecture-specific scheduling. TPUs are designed around batched parallel computation, which is exactly what MTP’s verification step does. Pixel TPU shows the highest speedups because the workload pattern aligns with the architecture’s strengths.

Where the RX 9060 XT lands: ~456 GB/s memory bandwidth, ~26 TFLOPS FP16. Bandwidth-to-compute ratio: ~17.5 GB/TFLOP — slightly more compute-bound than A100 (~26 GB/TFLOP) but still memory-pressured at this model size. IQ3_M quantisation reduces bandwidth pressure further, which cuts both ways. A reasonable projection range is 1.5–2.0×, with 2.5× as the optimistic ceiling.

TurboQuant (Released 24/03/2026)

Google Research’s TurboQuant compresses KV cache to 3 bits with reported zero accuracy loss, using polar-coordinate quantisation combined with a Johnson-Lindenstrauss residual error correction. Tested on Gemma and Mistral. The performance claim is up to 8× speedup over unquantised keys on H100 hardware.

For the build documented here, a 3-bit KV cache would compress to roughly 37% of the current Q8_0 size:

That headroom matters. It would either allow 262K to run with --parallel 4 (multi-slot at full native context), or leave room for an MTP drafter alongside the main model, or both.

If Both Land in llama.cpp

Combined, MTP and TurboQuant on this hardware could plausibly mean:

Worth pausing on this projection range. Even at the conservative 1.5× lower bound, 262K context decodes at 40 tok/s — the same speed as today’s 64K production figure, but at four times the context. The 2.5× optimistic ceiling puts 262K at 67 tok/s, which as recently as 2024 was data-centre territory. None of this has been measured on this build yet. But the projections are conservative against the published chart, the underlying mechanism is well-understood, and 16 GB consumer AMD hardware sits closer to genuinely useful long-context inference than most observers — local-first or cloud-first — currently price into their thinking.

The 31B Dense Question

Look at the A100 numbers again: 1.5× for 26B MoE, 3.0× for 31B Dense — same hardware, doubled gain. The bigger dense model exposes more memory pressure for MTP to capture back. If that pattern holds on consumer AMD, a 24 GB AMD card running Gemma 4 31B Dense with both MTP and TurboQuant could see substantially higher speedups than the 26B build documented here.

The MoE architecture trades total parameters for active parameters; the 31B Dense uses every parameter on every token, with the reasoning capacity that implies. At 16 GB with IQ3_M and current quantisation, 31B Dense does not fit usefully. With TurboQuant’s KV cache compression and MTP’s speedup, it might.

The build process documented here would translate directly. The hardware is the constraining piece — a 24 GB AMD card would bring 31B Dense well within the practical envelope and let a properly resourced AMD builder map territory this article cannot reach. If anyone in the AMD Dev Group community has access to hardware in that class and would like to compare notes on a 31B build, that is the experiment that would close this loop.

The Twilight Zone Summary

None of the speculative column has been measured on this build. All of it is conditional on llama.cpp implementing MTP and TurboQuant on the HIP backend, which is itself conditional on continued ROCm investment from AMD. Worth following.

Related Reading

This article is part of an ongoing series on local AI infrastructure and enterprise technology. The companion essay series Logical Zombies at zombies.ie covers AI and enterprise technology from a practitioner perspective.

Newbie_Article_1_ver07 | 09/05/2026 | Des Donnelly | https://www.dd.ie/