Add DeepSeek V4 target-only mlx-serve backend

.github/temporary-pr254-extra.py

@@ -0,0 +1,55 @@
+post_start_begin = (
+    '    if (\n'
+    '        "private func refreshPostStartState(" in ours\n'
+)
+post_start_end_marker = '        kinds.add("post_start")\n'
+post_start_start = rebase.find(post_start_begin)
+if post_start_start < 0:
+    raise SystemExit("post-start resolver start anchor changed")
+post_start_end = rebase.find(post_start_end_marker, post_start_start)
+if post_start_end < 0:
+    raise SystemExit("post-start resolver end anchor changed")
+post_start_end += len(post_start_end_marker)
+post_start_replacement = '''    if (
+        "configuration: MTPLXAppConfiguration," in ours
+        and "lifecycleEpoch: Int" in ours
+        and "recoveryGeneration: Int?" in ours
+        and ") async -> Bool" in ours
+        and "configuration: MTPLXAppConfiguration" in theirs
+        and "daemonBackendKind(for: configuration) == .mtplx" in theirs
+        and "external mlx-serve ready" in theirs
+        and "startExternalMlxServeHealthWatchdog" in theirs
+    ):
+        text = ours
+        if not text.endswith("\\n"):
+            text += "\\n"
+        text += (
+            "        guard daemonBackendKind(for: configuration) == .mtplx else {\\n"
+            "            health = nil\\n"
+            "            capabilities = nil\\n"
+            "            sessions = nil\\n"
+            "            sessionBank = nil\\n"
+            "            settings = nil\\n"
+            "            pendingLiveSettings = nil\\n"
+            "            pendingLiveSettingsModel = nil\\n"
+            "            connectionState = .idle\\n"
+            "            await supervisor.logs.append(\\n"
+            "                \\\"external mlx-serve ready; MTPLX live controls and metrics are unavailable\\\",\\n"
+            "                stream: .system\\n"
+            "            )\\n"
+            "            guard daemonSessionIsCurrent(\\n"
+            "                lifecycleEpoch: lifecycleEpoch,\\n"
+            "                launchID: nil,\\n"
+            "                recoveryGeneration: recoveryGeneration\\n"
+            "            ) else { return false }\\n"
+            "            startExternalMlxServeHealthWatchdog()\\n"
+            "            return daemonSessionIsCurrent(\\n"
+            "                lifecycleEpoch: lifecycleEpoch,\\n"
+            "                launchID: nil,\\n"
+            "                recoveryGeneration: recoveryGeneration\\n"
+            "            )\\n"
+            "        }\\n"
+        )
+        kinds.add("post_start")
+'''
+rebase = rebase[:post_start_start] + post_start_replacement + rebase[post_start_end:]

.github/workflows/temporary-materialize-pr254.yml

@@ -0,0 +1,247 @@
+name: Temporary materialize PR 254 current-main candidate
+
+on:
+  push:
+    branches:
+      - codex/deepseek-v4-mlxserve-v26
+
+permissions:
+  contents: write
+
+concurrency:
+  group: temporary-materialize-pr254
+  cancel-in-progress: true
+
+jobs:
+  materialize:
+    runs-on: macos-15
+    timeout-minutes: 45
+    env:
+      GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: codex/deepseek-v4-mlxserve-v26
+          fetch-depth: 0
+
+      - name: Rebuild and validate one clean feature commit on current upstream main
+        shell: bash
+        run: |
+          set -euo pipefail
+          git config user.name "Philip John Basile"
+          git config user.email "PBasile@Basilecom.com"
+          git remote add upstream https://github.com/youssofal/MTPLX.git || true
+          git fetch upstream main
+          git fetch origin agent/pr254-rebase-inspect
+          git show origin/agent/pr254-rebase-inspect:.github/workflows/temporary-rebase-254.yml > /tmp/pr254-source.yml
+
+          python - <<'PY'
+          from pathlib import Path
+
+          lines = Path("/tmp/pr254-source.yml").read_text(encoding="utf-8").splitlines()
+          blocks: list[str] = []
+          index = 0
+          while index < len(lines):
+              if lines[index] == "        run: |":
+                  index += 1
+                  block: list[str] = []
+                  while index < len(lines):
+                      current = lines[index]
+                      if current and len(current) - len(current.lstrip()) <= 8:
+                          break
+                      if current.startswith("          "):
+                          block.append(current[10:])
+                      elif not current:
+                          block.append("")
+                      else:
+                          raise SystemExit(f"unexpected workflow indentation: {current!r}")
+                      index += 1
+                  blocks.append("\n".join(block))
+                  continue
+              index += 1
+          if len(blocks) < 3:
+              raise SystemExit(f"expected at least three guarded run blocks, found {len(blocks)}")
+
+          rebase = blocks[0]
+          old_expected = '  test "$conflicts" = "mtplx/hf_loader.py"'
+          new_expected = (
+              "  expected=$'apps/MTPLXApp/Sources/MTPLXAppCore/Services/"
+              "DaemonSupervisor.swift\\napps/MTPLXApp/Sources/MTPLXAppCore/Stores/"
+              "MTPLXBackendStore.swift\\nmtplx/hf_loader.py'\n"
+              "  test \"$conflicts\" = \"$expected\""
+          )
+          if rebase.count(old_expected) != 1:
+              raise SystemExit("hf_loader conflict expectation anchor changed")
+          rebase = rebase.replace(old_expected, new_expected, 1)
+
+          old_continue = (
+              "  git add mtplx/hf_loader.py\n"
+              "  set +e\n"
+              "  GIT_EDITOR=true git rebase --continue\n"
+              "  rc=$?\n"
+              "  set -e"
+          )
+          new_continue = (
+              "  git add mtplx/hf_loader.py\n"
+              "  # The two Swift paths remain unmerged in this same rebase stop.\n"
+              "  rc=1"
+          )
+          if rebase.count(old_continue) != 1:
+              raise SystemExit("hf_loader continuation anchor changed")
+          rebase = rebase.replace(old_continue, new_continue, 1)
+
+          def replace_resolver(start_marker: str, end_marker: str, replacement: str, label: str) -> None:
+              global rebase
+              start = rebase.find(start_marker)
+              if start < 0:
+                  raise SystemExit(f"{label} resolver start anchor changed")
+              end = rebase.find(end_marker, start)
+              if end < 0:
+                  raise SystemExit(f"{label} resolver end anchor changed")
+              end += len(end_marker)
+              rebase = rebase[:start] + replacement + rebase[end:]
+
+          replace_resolver(
+              '    if (\n        "private func waitForHealth(" in ours\n',
+              '        kinds.add("wait_for_health")\n',
+              (
+                  '    if (\n'
+                  '        ours.strip() == ") async throws -> HealthPayload {"\n'
+                  '        and ") async throws -> HealthPayload?" in theirs\n'
+                  '        and "waitForExternalMlxServeHealth" in theirs\n'
+                  '        and "let client = MTPLXAPIClient" in theirs\n'
+                  '    ):\n'
+                  '        text = theirs\n'
+                  '        kinds.add("wait_for_health")\n'
+              ),
+              "waitForHealth",
+          )
+
+          replace_resolver(
+              (
+                  '    if "public func refreshStaticState(" in ours '
+                  'and "supportsMTPLXLiveControls" in theirs:\n'
+              ),
+              '        kinds.add("refresh_static")\n',
+              (
+                  '    if (\n'
+                  '        "public func refreshStaticState(" in ours\n'
+                  '        and "public func refreshStaticState()" in theirs\n'
+                  '        and "guard supportsMTPLXLiveControls else { return }" in theirs\n'
+                  '        and "let client = apiClient" not in ours\n'
+                  '    ):\n'
+                  '        text = ours\n'
+                  '        if not text.endswith("\\n"):\n'
+                  '            text += "\\n"\n'
+                  '        text += (\n'
+                  '            "        guard supportsMTPLXLiveControls else {\\n"\n'
+                  '            "            health = nil\\n"\n'
+                  '            "            capabilities = nil\\n"\n'
+                  '            "            sessions = nil\\n"\n'
+                  '            "            prefillHistory = nil\\n"\n'
+                  '            "            models = nil\\n"\n'
+                  '            "            return\\n"\n'
+                  '            "        }\\n"\n'
+                  '        )\n'
+                  '        kinds.add("refresh_static")\n'
+              ),
+              "refreshStaticState",
+          )
+
+          replace_resolver(
+              (
+                  '    if (\n'
+                  '        "private func flushFreshLaunchLiveOnlySettingsIfNeeded(" in ours\n'
+              ),
+              '        kinds.add("flush_fresh")\n',
+              (
+                  '    if (\n'
+                  '        "private func flushFreshLaunchLiveOnlySettingsIfNeeded(" in ours\n'
+                  '        and "private func flushFreshLaunchLiveOnlySettingsIfNeeded()" in theirs\n'
+                  '        and "guard supportsMTPLXLiveControls else {" in theirs\n'
+                  '        and "guard let pending = pendingLiveSettings" not in ours\n'
+                  '    ):\n'
+                  '        text = ours\n'
+                  '        if not text.endswith("\\n"):\n'
+                  '            text += "\\n"\n'
+                  '        text += (\n'
+                  '            "        guard supportsMTPLXLiveControls else {\\n"\n'
+                  '            "            pendingLiveSettings = nil\\n"\n'
+                  '            "            pendingLiveSettingsModel = nil\\n"\n'
+                  '            "            settings = nil\\n"\n'
+                  '            "            return\\n"\n'
+                  '            "        }\\n"\n'
+                  '        )\n'
+                  '        kinds.add("flush_fresh")\n'
+              ),
+              "flushFreshLaunchLiveOnlySettingsIfNeeded",
+          )
+
+          extra_patch = Path(".github/temporary-pr254-extra.py")
+          if extra_patch.exists():
+              namespace = {"rebase": rebase}
+              exec(compile(extra_patch.read_text(encoding="utf-8"), str(extra_patch), "exec"), namespace)
+              rebase = namespace["rebase"]
+
+          blocks[0] = rebase
+          Path("/tmp/pr254-rebase.sh").write_text(blocks[0] + "\n", encoding="utf-8")
+          Path("/tmp/pr254-squash.sh").write_text(blocks[1] + "\n", encoding="utf-8")
+          Path("/tmp/pr254-validate.sh").write_text(blocks[2] + "\n", encoding="utf-8")
+          PY
+
+          cat > /tmp/pr254-run.sh <<'SH'
+          set -euo pipefail
+          git reset --hard d079df37916f646f0a5b34343f317a03aa3a2e07
+          bash -n /tmp/pr254-rebase.sh
+          bash -n /tmp/pr254-squash.sh
+          bash -n /tmp/pr254-validate.sh
+          bash /tmp/pr254-rebase.sh
+          rm -f \
+            .github/workflows/temporary-materialize-pr254.yml \
+            .github/workflows/temporary-rebase-254.yml \
+            .github/workflows/temporary-execute-rebase-254-macos14.yml \
+            .github/workflows/temporary-execute-rebase-254-quote-fix.yml \
+            .github/pr254-wrapper-template.yml \
+            .github/temporary-rebase-254-trigger.txt \
+            .github/temporary-pr254-extra.py
+          bash /tmp/pr254-squash.sh
+          test "$(git rev-list --count upstream/main..HEAD)" = "1"
+          test "$(git rev-parse HEAD^)" = "$(git rev-parse upstream/main)"
+          test -z "$(git diff --name-only upstream/main...HEAD -- '.github/workflows/temporary-*' '.github/pr254-wrapper-template.yml' '.github/temporary-*')"
+          git show -s --format=%B HEAD | grep -F "Signed-off-by: Philip John Basile <PBasile@Basilecom.com>"
+          git diff --check upstream/main...HEAD
+          bash /tmp/pr254-validate.sh
+          SH
+
+          set +e
+          bash /tmp/pr254-run.sh > /tmp/pr254-run.log 2>&1
+          status=$?
+          set -e
+          cat /tmp/pr254-run.log
+
+          if [ "$status" -ne 0 ]; then
+            git rebase --abort >/dev/null 2>&1 || true
+            git reset --hard upstream/main
+            mkdir -p diagnostics
+            cp /tmp/pr254-run.log diagnostics/pr254-current-main-failure.log
+            git add diagnostics/pr254-current-main-failure.log
+            git commit -m "diagnostics: capture PR 254 current-main failure"
+            git push \
+              "https://x-access-token:${GH_TOKEN}@github.com/PhilipJohnBasile/MTPLX.git" \
+              HEAD:agent/pr254-failure-log --force
+            exit "$status"
+          fi
+
+          candidate="$(git rev-parse HEAD)"
+          git push \
+            "https://x-access-token:${GH_TOKEN}@github.com/PhilipJohnBasile/MTPLX.git" \
+            HEAD:sync/pr254-rebased-inspect --force
+          {
+            echo "## PR 254 current-main candidate"
+            echo ""
+            echo "- upstream main: \`$(git rev-parse upstream/main)\`"
+            echo "- validated candidate: \`$candidate\`"
+            echo "- commits ahead: 1"
+            echo "- temporary repair files: absent"
+            echo "- focused Python, package, hygiene, full Swift test, and release build: passed"
+          } >> "$GITHUB_STEP_SUMMARY"

README.md

@@ -165,6 +165,30 @@ comfortable. MTPLX defaults Laguna to a 32,768-token context
 and response cap, and checks larger explicit server contexts against the active
 Metal memory cap.
 
+[DeepSeek-V4-Flash-0731 MLX M5 Max Target-Only](https://huggingface.co/philipjohnbasile/DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly)
+uses a separate, experimental external runtime route. MTPLX pins the public
+artifact to `ac33e4f3ca3546e6cec104558d42161e15814e33`, admits the exact 44
+weight shards and required sidecars, then delegates serving to a separately
+installed `mlx-serve` executable. This is target-only AR — it has no MTP or
+DSpark weights — and it is not a native MTPLX backend:
+
+```bash
+mtplx pull philipjohnbasile/DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly
+
+MTPLX_MLX_SERVE_BIN=/path/to/mlx-serve \
+mtplx serve \
+  --model philipjohnbasile/DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly \
+  --no-mtp --host 127.0.0.1 --port 8000 --yes
+```
+
+The route requires a 128 GB Apple Silicon Mac, defaults to an 8,192-token
+context, disables PLD, decode-attention quantization, and vision, and preserves
+the external runtime's memory preflight. MTPLX clears ambient `MLX_SERVE_*`
+settings and launches with `MLX_SERVE_WIRED=fit` plus a 256 MB cache limit;
+set `MTPLX_DSV4_WIRED` only to make an explicit override. Representative
+streaming performance is unapproved, so neither this integration nor its dry
+run output makes a throughput claim.
+
 ## What MTPLX is not
 
 - Not an external-drafter system. The drafter is the target model's own MTP heads.