fix(earn): collapse catalog variants to one model per family by Gajesh2007 · Pull Request #417 · Layr-Labs/d-inference

Gajesh2007 · 2026-06-20T03:29:23Z

Summary

Follow-up to #414. The earnings calculator was listing every catalog build of a model as a separate row. The live catalog returns four entries:

gpt-oss-20b → "GPT-OSS 20B"
gemma-4-26b → "Gemma 4 26B" (active)
gemma-4-26b-qat-4bit → "Gemma 4 26B" (beta, duplicate name)
gemma-4-26b-8bit → "Gemma 4 26B 8-bit (rollback)"

…so the picker showed 4 rows (two identical "Gemma 4 26B", an 8-bit rollback, and GPT-OSS). It should show the two real models: Gemma 4 26B and GPT-OSS 20B.

Fix

Dedupe the catalog by a normalized base-model key (baseModelKey strips quantization / build suffixes like -qat-4bit, -8bit, -rollback, -bf16, …) and keep the most canonical build per family (variantPenalty favors clean names and shorter base ids over quant/rollback builds).

No hardcoded model list — derived from the live catalog, so a genuinely new model still appears; only quant/rollback variants of the same base collapse.
Mirrored across console-ui/src/app/earn/calc.ts (dedupeModelVariants, used in buildCatalogModels) and landing/earn-calculator.js.
Added unit tests: baseModelKey suffix stripping, dedupeModelVariants → one entry per base, and buildCatalogModels → exactly ["GPT-OSS 20B", "Gemma 4 26B"].

Before / after

	Models shown
Before	Gemma 4 26B · Gemma 4 26B · Gemma 4 26B 8-bit (rollback) · GPT-OSS 20B
After	Gemma 4 26B · GPT-OSS 20B

Verification

npx vitest run: ✅ (earn suite 6/6, incl. 3 new dedupe tests)
npx eslint src/app/earn/calc.ts: ✅ 0 errors
node --check landing/earn-calculator.js: ✅

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

The calculator listed every catalog build — gemma-4-26b, gemma-4-26b-qat-4bit and gemma-4-26b-8bit (rollback) all showed as separate rows alongside gpt-oss-20b (4 entries). Dedupe by a normalized base-model key (strips quant/rollback suffixes) and keep the canonical build, so only the two real models show: Gemma 4 26B and GPT-OSS 20B. No hardcoded model list — derived from the live catalog. Mirrored across console /earn and landing; added unit tests.

vercel · 2026-06-20T03:29:27Z

@Gajesh2007 is attempting to deploy a commit to the EigenLabs Team on Vercel.

A member of the Team first needs to authorize it.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dc8bed2813

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-06-20T03:42:02Z

+    const text = `${m.display_name || ""} ${m.id || ""}`.toLowerCase();
+    let p = 0;
+    if (/\(|rollback|preview|\brc\b/.test(text)) p += 100;
+    if (/qat|int4|int8|fp16|bf16|mxfp4|nf4|\d\s*-?bit/.test(text)) p += 10;


Prefer alias primary builds when collapsing variants

When the landing calculator consumes /v1/models/catalog during an alias takeover, the catalog can contain both the public-name previous build (gemma-4-26b) and the quantized desired build (gemma-4-26b-qat-4bit), while the coordinator routes the public alias to the desired build. This penalty makes the desired build lose solely because its id/name contains qat/4bit, so the single displayed Gemma row can use the old build's size and prices even though requests are routed to the new build; use the alias primary_build metadata or otherwise prefer the routed/desired build before penalizing quant suffixes.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-20T03:42:02Z

 export function buildCatalogModels(models: Model[], pricing: PricingResponse | null): CatalogModel[] {
  const prices = buildPricingLookup(pricing);
-  return models
+  return dedupeModelVariants(models)


Preserve lower-memory variants during dedupe

Because variants are collapsed before the page applies its RAM eligibility filter, a clean/base row with a higher min_ram_gb or larger size_gb can replace a quantized sibling that would fit on smaller hardware. For those RAM selections, rankedModels later filters out the retained base row and the fitting variant is already gone, so users can see no compatible model even though the catalog has one; the dedupe needs to account for eligibility or retain the lowest-memory representative when variants differ in fit.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-20T03:42:02Z

+ */
+export function baseModelKey(id: string): string {
+  let k = id.toLowerCase().trim();
+  const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;


Strip fp8 variants in base model keys

The alias migration paths in this repo use previous builds ending in -fp8, but this suffix list strips -qat-4bit while leaving -fp8 intact. In a catalog containing an FP8 previous build and a QAT desired build for the same family, they get different base keys and both remain in the earnings picker, so the original duplicate-variant problem still appears for that rollout shape; include FP8 (and related fp quant suffixes) in the normalized key.

Useful? React with 👍 / 👎.

ethenotethan

Automated Code Review — Layr-Labs/d-inference#

Verdict: REQUEST_CHANGES

Security — ✅ No issues found

Performance — 2 finding(s) (2 blocking)

🟡 [MEDIUM] console-ui/src/app/earn/calc.ts:245-252 — Map lookup in tight loop for model deduplication
- Suggestion: Consider pre-sorting models by penalty score to avoid repeated Map.get() calls, or use a single pass with array sorting if the number of models is small
🟡 [MEDIUM] landing/earn-calculator.js:133-140 — Duplicate Map lookup pattern in dedupeModelVariants
- Suggestion: Same optimization as TypeScript version - consider pre-sorting or single-pass approach

Type_diligence — ✅ No issues found

Additive_complexity — 3 finding(s) (1 blocking)

🔵 [INFO] console-ui/src/app/earn/calc.ts:229 — variantPenalty function could be simplified with a single regex
- Suggestion: Combine the two regex checks into one pattern or use a simple scoring array instead of complex penalty calculation
🟡 [MEDIUM] landing/earn-calculator.js:118-141 — Exact duplication of baseModelKey, variantPenalty, and dedupeModelVariants logic
- Suggestion: Extract this logic into a shared utility module or consider if the landing page really needs this complexity vs a simpler hardcoded list
🔵 [INFO] console-ui/src/app/earn/calc.ts:218-227 — Complex regex pattern with many quantization suffixes may be overkill
- Suggestion: Consider if a simpler approach like checking for common suffixes (-4bit, -8bit, -rollback) would suffice for the current use case

5 finding(s) total, 3 blocking. Verdict: REQUEST_CHANGES.

🤖 Automated review by Centaur · DAR-186

ethenotethan · 2026-06-20T05:59:16Z

+export function dedupeModelVariants(models: Model[]): Model[] {
+  const byBase = new Map<string, Model>();
+  for (const m of models) {
+    const key = baseModelKey(m.id);
+    const cur = byBase.get(key);
+    if (!cur || variantPenalty(m) < variantPenalty(cur)) byBase.set(key, m);
+  }
+  return [...byBase.values()];


🟡 [MEDIUM] ⚡ Map lookup in tight loop for model deduplication

💡 Suggestion: Consider pre-sorting models by penalty score to avoid repeated Map.get() calls, or use a single pass with array sorting if the number of models is small

📊 Score: 2×4 = 8 · Category: Inefficient data structures

ethenotethan · 2026-06-20T05:59:17Z

+  return k;
+}
+
+/** Lower is more canonical — prefer clean names / base ids over quant/rollback builds. */


🔵 [INFO] 🧩 variantPenalty function could be simplified with a single regex

💡 Suggestion: Combine the two regex checks into one pattern or use a simple scoring array instead of complex penalty calculation

📊 Score: 2×3 = 6 · Category: over-abstraction

ethenotethan · 2026-06-20T05:59:17Z

+  function baseModelKey(id) {
+    let k = String(id || "").toLowerCase().trim();
+    const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+    let prev = "";
+    while (k !== prev) { prev = k; k = k.replace(suffix, ""); }
+    return k;
+  }
+  function variantPenalty(m) {
+    const text = `${m.display_name || ""} ${m.id || ""}`.toLowerCase();
+    let p = 0;
+    if (/\(|rollback|preview|\brc\b/.test(text)) p += 100;
+    if (/qat|int4|int8|fp16|bf16|mxfp4|nf4|\d\s*-?bit/.test(text)) p += 10;
+    p += String(m.id || "").length * 0.01;
+    return p;
+  }
+  function dedupeModelVariants(models) {
+    const byBase = new Map();
+    for (const m of models) {
+      const key = baseModelKey(m.id);
+      const cur = byBase.get(key);
+      if (!cur || variantPenalty(m) < variantPenalty(cur)) byBase.set(key, m);
+    }
+    return [...byBase.values()];
+  }


🟡 [MEDIUM] 🧩 Exact duplication of baseModelKey, variantPenalty, and dedupeModelVariants logic

💡 Suggestion: Extract this logic into a shared utility module or consider if the landing page really needs this complexity vs a simpler hardcoded list

📊 Score: 3×4 = 12 · Category: duplicate logic

ethenotethan · 2026-06-20T05:59:17Z

+export function baseModelKey(id: string): string {
+  let k = id.toLowerCase().trim();
+  const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+  let prev = "";
+  while (k !== prev) {
+    prev = k;
+    k = k.replace(suffix, "");
+  }
+  return k;
+}


🔵 [INFO] 🧩 Complex regex pattern with many quantization suffixes may be overkill

💡 Suggestion: Consider if a simpler approach like checking for common suffixes (-4bit, -8bit, -rollback) would suffice for the current use case

📊 Score: 2×2 = 4 · Category: over-configuration

ethenotethan

Automated Code Review — Layr-Labs/d-inference#

Verdict: REQUEST_CHANGES

Security — ✅ No issues found

Performance — 4 finding(s) (1 blocking)

🟡 [MEDIUM] console-ui/src/app/earn/calc.ts:246-250 — O(n²) nested loop in dedupeModelVariants due to Map.get() calls inside loop
- Suggestion: The current implementation is already efficient with O(n) complexity using Map operations. No optimization needed.
🔵 [INFO] landing/earn-calculator.js:133-140 — Redundant regex compilation and string operations in baseModelKey loop
- Suggestion: Pre-compile the regex outside the function and consider caching results for repeated calls with same input
🔵 [INFO] console-ui/src/app/earn/calc.ts:218-227 — Regex compilation happens on every baseModelKey call
- Suggestion: Move regex compilation outside function scope: const SUFFIX_REGEX = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
🔵 [INFO] landing/earn-calculator.js:118-122 — Regex compilation happens on every baseModelKey call
- Suggestion: Move regex compilation outside function scope to avoid recompiling on each call

Type_diligence — ✅ No issues found

Additive_complexity — 2 finding(s) (1 blocking)

🔵 [INFO] console-ui/src/app/earn/calc.ts:219-227 — Complex regex pattern for suffix stripping could be simplified
- Suggestion: Consider using a simple array of known suffixes and string.endsWith() checks instead of a complex regex with while loop
🟡 [MEDIUM] landing/earn-calculator.js:118-140 — Exact duplication of baseModelKey, variantPenalty, and dedupeModelVariants logic
- Suggestion: Extract this logic into a shared utility module that both console-ui and landing can import, or generate the landing code from the TypeScript source

6 finding(s) total, 2 blocking. Verdict: REQUEST_CHANGES.

🤖 Automated review by Centaur · DAR-186

ethenotethan · 2026-06-20T06:04:40Z

+  const byBase = new Map<string, Model>();
+  for (const m of models) {
+    const key = baseModelKey(m.id);
+    const cur = byBase.get(key);
+    if (!cur || variantPenalty(m) < variantPenalty(cur)) byBase.set(key, m);


🟡 [MEDIUM] ⚡ O(n²) nested loop in dedupeModelVariants due to Map.get() calls inside loop

💡 Suggestion: The current implementation is already efficient with O(n) complexity using Map operations. No optimization needed.

📊 Score: 2×4 = 8 · Category: Inefficient data structures

ethenotethan · 2026-06-20T06:04:40Z

+  function dedupeModelVariants(models) {
+    const byBase = new Map();
+    for (const m of models) {
+      const key = baseModelKey(m.id);
+      const cur = byBase.get(key);
+      if (!cur || variantPenalty(m) < variantPenalty(cur)) byBase.set(key, m);
+    }
+    return [...byBase.values()];


🔵 [INFO] ⚡ Redundant regex compilation and string operations in baseModelKey loop

💡 Suggestion: Pre-compile the regex outside the function and consider caching results for repeated calls with same input

📊 Score: 2×3 = 6 · Category: Repeated work

ethenotethan · 2026-06-20T06:04:40Z

+export function baseModelKey(id: string): string {
+  let k = id.toLowerCase().trim();
+  const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+  let prev = "";
+  while (k !== prev) {
+    prev = k;
+    k = k.replace(suffix, "");
+  }
+  return k;
+}


🔵 [INFO] ⚡ Regex compilation happens on every baseModelKey call

💡 Suggestion: Move regex compilation outside function scope: const SUFFIX_REGEX = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;

📊 Score: 1×2 = 2 · Category: Repeated work

ethenotethan · 2026-06-20T06:04:40Z

+  function baseModelKey(id) {
+    let k = String(id || "").toLowerCase().trim();
+    const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+    let prev = "";
+    while (k !== prev) { prev = k; k = k.replace(suffix, ""); }


🔵 [INFO] ⚡ Regex compilation happens on every baseModelKey call

💡 Suggestion: Move regex compilation outside function scope to avoid recompiling on each call

📊 Score: 1×2 = 2 · Category: Repeated work

ethenotethan · 2026-06-20T06:04:40Z

+  let k = id.toLowerCase().trim();
+  const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+  let prev = "";
+  while (k !== prev) {
+    prev = k;
+    k = k.replace(suffix, "");
+  }
+  return k;
+}


🔵 [INFO] 🧩 Complex regex pattern for suffix stripping could be simplified

💡 Suggestion: Consider using a simple array of known suffixes and string.endsWith() checks instead of a complex regex with while loop

📊 Score: 2×3 = 6 · Category: over-abstraction

ethenotethan · 2026-06-20T06:04:40Z

+  function baseModelKey(id) {
+    let k = String(id || "").toLowerCase().trim();
+    const suffix = /-(qat|q4|q8|int4|int8|4bit|8bit|4-bit|8-bit|bf16|fp16|mxfp4|nf4|gguf|rollback|preview|beta|rc\d*)$/;
+    let prev = "";
+    while (k !== prev) { prev = k; k = k.replace(suffix, ""); }
+    return k;
+  }
+  function variantPenalty(m) {
+    const text = `${m.display_name || ""} ${m.id || ""}`.toLowerCase();
+    let p = 0;
+    if (/\(|rollback|preview|\brc\b/.test(text)) p += 100;
+    if (/qat|int4|int8|fp16|bf16|mxfp4|nf4|\d\s*-?bit/.test(text)) p += 10;
+    p += String(m.id || "").length * 0.01;
+    return p;
+  }
+  function dedupeModelVariants(models) {
+    const byBase = new Map();
+    for (const m of models) {
+      const key = baseModelKey(m.id);
+      const cur = byBase.get(key);
+      if (!cur || variantPenalty(m) < variantPenalty(cur)) byBase.set(key, m);
+    }
+    return [...byBase.values()];


🟡 [MEDIUM] 🧩 Exact duplication of baseModelKey, variantPenalty, and dedupeModelVariants logic

💡 Suggestion: Extract this logic into a shared utility module that both console-ui and landing can import, or generate the landing code from the TypeScript source

📊 Score: 3×4 = 12 · Category: duplicate logic

Gajesh2007 had a problem deploying to benchmarks June 20, 2026 03:29 — with GitHub Actions Failure

Merge branch 'master' into fix/earn-dedupe-model-variants

c552847

Gajesh2007 had a problem deploying to benchmarks June 20, 2026 03:30 — with GitHub Actions Failure

Gajesh2007 merged commit 916de68 into Layr-Labs:master Jun 20, 2026
4 of 11 checks passed

chatgpt-codex-connector Bot reviewed Jun 20, 2026

View reviewed changes

ethenotethan requested changes Jun 20, 2026

View reviewed changes

Uh oh!

Conversation

Gajesh2007 commented Jun 20, 2026 • edited by blacksmith-sh Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Fix

Before / after

Verification

Uh oh!

vercel Bot commented Jun 20, 2026

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jun 20, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 20, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 20, 2026

Choose a reason for hiding this comment

Uh oh!

ethenotethan left a comment

Choose a reason for hiding this comment

Automated Code Review — Layr-Labs/d-inference#

Security — ✅ No issues found

Performance — 2 finding(s) (2 blocking)

Type_diligence — ✅ No issues found

Additive_complexity — 3 finding(s) (1 blocking)

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ethenotethan left a comment

Choose a reason for hiding this comment

Automated Code Review — Layr-Labs/d-inference#

Security — ✅ No issues found

Performance — 4 finding(s) (1 blocking)

Type_diligence — ✅ No issues found

Additive_complexity — 2 finding(s) (1 blocking)

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Gajesh2007 commented Jun 20, 2026 •

edited by blacksmith-sh Bot

Loading