Log cheapest RunPod search

Author: cderinbogaz

experiments/002-kimi-k26-gguf-q2/2026-04-25-runpod-cheapest-search.md

@@ -0,0 +1,106 @@
+# 2026-04-25 RunPod Cheapest Search for Kimi K2.6 GGUF Q2
+
+## Goal
+
+Find the cheapest currently available RunPod topology that can plausibly run `unsloth/Kimi-K2.6-GGUF:UD-Q2_K_XL`, then start it and benchmark if the runtime becomes reachable.
+
+## Selection Rule
+
+Use even-GPU shapes only, and prefer the lowest-cost topology that clears the rough aggregate-VRAM floor for the `340 GB` GGUF artifact.
+
+## Live Cheap-First Search
+
+### 1. 8x A40
+
+- Status in inventory: available, low stock
+- Result: `HTTP 500 There are no instances currently available`
+- Outcome: could not allocate
+
+### 2. 8x RTX A6000
+
+- Status in inventory: advertised in several regions, but without a strong stock signal
+- Result: `HTTP 500 There are no instances currently available`
+- Outcome: could not allocate
+
+### 3. 2x MI300X
+
+| Field | Value |
+| --- | --- |
+| Pod ID | `k9p5qwst0txevv` |
+| Cost | `$3.98/hr` |
+| Machine ID | `j03rnq2tcsxu` |
+| Result | allocated |
+
+Observed behavior:
+
+- `desiredStatus: RUNNING`
+- `uptimeSeconds: 0`
+- no public routing
+- SSH never reached `ready`
+
+Outcome: cheapest allocatable option, but dead host.
+
+### 4. 4x H100 80GB
+
+- Status in inventory: high-level stock existed, but not in the REST-allowed regions that were tested
+- Result: `HTTP 500 There are no instances currently available`
+- Outcome: could not allocate
+
+### 5. 4x H100 NVL
+
+| Field | Value |
+| --- | --- |
+| Pod ID | `j6q4iu80tj922e` |
+| Cost | `$10.36/hr` |
+| Machine ID | `o7h99o28jtin` |
+| Result | allocated |
+
+Observed behavior:
+
+- new machine ID, unlike the recycled RTX PRO 6000 community host
+- SSH metadata appeared at `38.143.35.131:12908`
+- direct SSH still returned `Connection refused`
+- `uptimeSeconds` remained `0`
+
+Outcome: allocates, but still dead before runtime.
+
+### 6. 4x H200
+
+- Status in inventory: low stock in some regions
+- Result: `HTTP 500 There are no instances currently available`
+- Outcome: could not allocate
+
+### 7. 4x RTX PRO 6000 Secure
+
+| Field | Value |
+| --- | --- |
+| Pod ID | `c3j21r1pd9wpa2` |
+| Cost | `$7.56/hr` |
+| Machine ID | `67fbuhb2qnz1` |
+| Datacenter | `EUR-IS-1` |
+| Result | allocated |
+
+Observed behavior:
+
+- first Secure RTX PRO 6000 host tested, so this was a different pool than the recycled dead community host
+- SSH metadata appeared at `157.157.221.30:52123`
+- direct SSH still returned `Connection refused`
+- `uptimeSeconds` remained `0`
+
+Outcome: cheapest allocatable NVIDIA path found today, but still dead before runtime.
+
+## Practical Conclusion
+
+As of 2026-04-25, the cheapest allocatable RunPod topology found for this model was:
+
+- `2x MI300X` at `$3.98/hr`
+
+The cheapest allocatable NVIDIA topology found was:
+
+- `4x RTX PRO 6000 Secure` at `$7.56/hr`
+
+Neither became reachable enough to run inference, so no successful benchmark could be produced from the cheap-first search.
+
+## Cleanup
+
+Every failed pod from this search was deleted after its readiness window.

experiments/002-kimi-k26-gguf-q2/README.md

@@ -92,6 +92,28 @@ Outcome:
 
 These attempts confirm that the current RunPod `4x RTX PRO 6000` allocator is also returning a non-ready host for this workflow.
 
+## 2026-04-25 Cheapest-Available Search
+
+An explicit cheapest-first search was run on 2026-04-25 against the live RunPod inventory for the `UD-Q2_K_XL` GGUF path. The target requirement was the smallest practical even-GPU topology with enough aggregate VRAM to plausibly serve the 340 GB artifact.
+
+Observed order from the live market:
+
+| Candidate | Result |
+| --- | --- |
+| `8x A40` | cheapest likely fit, but no allocatable instances |
+| `8x RTX A6000` | no allocatable instances |
+| `2x MI300X` | allocated at `$3.98/hr`, but host never transitioned into a live runtime |
+| `4x H100 80GB` | no allocatable instances in REST-allowed H100 regions |
+| `4x H100 NVL` | allocated at `$10.36/hr`, but host never transitioned into a live runtime |
+| `4x H200` | no allocatable instances |
+| `4x RTX PRO 6000 Secure` | allocated at `$7.56/hr`, new host, but still never transitioned into a live runtime |
+
+Conclusion:
+
+- The cheapest allocatable shape today was `2x MI300X` at `$3.98/hr`.
+- The cheapest allocatable NVIDIA shape today was `4x RTX PRO 6000 Secure` at `$7.56/hr`.
+- None of the allocatable shapes actually became reachable enough to run inference, so there is still no successful cheap RunPod baseline for this model on this date.
+
 ## Launch Runbook
 
 Capacity probe with a disposable pod volume: