Merge branch 'upstream' into concedo_experimental

# Conflicts: # .devops/cann.Dockerfile # .devops/cpu.Dockerfile # .devops/cuda.Dockerfile # .devops/intel.Dockerfile # .devops/llama-cli-cann.Dockerfile # .devops/musa.Dockerfile # .devops/openvino.Dockerfile # .devops/rocm.Dockerfile # .devops/s390x.Dockerfile # .devops/vulkan.Dockerfile # .github/ISSUE_TEMPLATE/011-bug-results.yml # .github/ISSUE_TEMPLATE/019-bug-misc.yml # .github/workflows/build-and-test-snapdragon.yml # .github/workflows/docker.yml # .github/workflows/server-self-hosted.yml # .github/workflows/ui-ci.yml # .pi/gg/SYSTEM.md # README.md # common/arg.cpp # docs/backend/SYCL.md # docs/backend/snapdragon/CMakeUserPresets.json # docs/backend/snapdragon/README.md # docs/speculative.md # examples/save-load-state/save-load-state.cpp # ggml/src/ggml-hexagon/ggml-hexagon.cpp # ggml/src/ggml-hexagon/htp/CMakeLists.txt # ggml/src/ggml-hexagon/htp/htp-ctx.h # ggml/src/ggml-hexagon/htp/htp-ops.h # ggml/src/ggml-hexagon/htp/main.c # ggml/src/ggml-hexagon/htp/rope-ops.c # ggml/src/ggml-hexagon/htp/unary-ops.c # ggml/src/ggml-opencl/CMakeLists.txt # ggml/src/ggml-opencl/ggml-opencl.cpp # ggml/src/ggml-opencl/kernels/cvt.cl # ggml/src/ggml-sycl/ggml-sycl.cpp # ggml/src/ggml-webgpu/ggml-webgpu.cpp # ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl # tools/cli/README.md # tools/server/README.md
2026-09-19 09:15:18 +02:00 · 2026-05-20 18:48:34 +08:00
parent 643862ac7d 585080d310
commit 7d987af23a
77 changed files with 1291 additions and 1048 deletions
@@ -149,6 +149,8 @@ class TaskState:
    t_gen_ms: Optional[float] = None
    reasoning_content: Optional[str] = None
    server_name: Optional[str] = None
+    chunk_idx: int = 0
+    problem_idx: int = 0


 class EvalState:
@@ -233,7 +235,9 @@ class EvalState:
        tps_gen: Optional[float] = None,
        t_gen_ms: Optional[float] = None,
        reasoning_content: Optional[str] = None,
-        server_name: Optional[str] = None
+        server_name: Optional[str] = None,
+        chunk_idx: int = 0,
+        problem_idx: int = 0,
    ):
        with self._lock:
            if "cases" not in self.task_states:
@@ -252,7 +256,9 @@ class EvalState:
                "tps_gen": tps_gen,
                "t_gen_ms": t_gen_ms,
                "reasoning_content": reasoning_content,
-                "server_name": server_name
+                "server_name": server_name,
+                "chunk_idx": chunk_idx,
+                "problem_idx": problem_idx,
            }

            self.correct = sum(1 for c in self.task_states.get("cases", {}).values() if c.get("correct", False))
@@ -289,6 +295,9 @@ class EvalState:
            all_cases = {}
            for i, task_id in tasks_to_save:
                question_text, prompt, expected = self.get_case(i)
+                # Extract chunk_idx from task_id for pending cases
+                _parts = task_id.rsplit("_", 2)
+                _chunk_idx = int(_parts[-2]) if len(_parts) >= 3 else 0
                if task_id in self.task_states.get("cases", {}):
                    all_cases[task_id] = self.task_states["cases"][task_id]
                else:
@@ -306,7 +315,9 @@ class EvalState:
                        "tps_gen": None,
                        "t_gen_ms": None,
                        "reasoning_content": None,
-                        "server_name": None
+                        "server_name": None,
+                        "chunk_idx": _chunk_idx,
+                        "problem_idx": i,
                    }

            ci_lower, ci_upper = self.accuracy_ci()
@@ -382,11 +393,12 @@ class EvalState:
            grader_log_str = self._escape_html(json.dumps(grader_log, indent=2))
            escaped_server = self._escape_html(server_name)

+            answer_class = status_class if status == "ok" else ""
            rows.append(f"""<tr class="task-row" onclick="toggleDetails('{task_id}')">
                <td>{task_id}</td>
                <td class="{status_class}">{status_text}</td>
                <td>{self._escape_html(expected)}</td>
-                <td>{self._escape_html(answer)}</td>
+                <td class="{answer_class}">{self._escape_html(answer)}</td>
                <td>{tokens_str}</td>
                <td>{tps_str}</td>
                <td>{t_gen_str}</td>
@@ -405,6 +417,53 @@ class EvalState:

        rows_html = "\n".join(rows)

+        # ---- per-problem summary table ----
+        problem_groups: Dict[int, List[Dict[str, Any]]] = {}
+        for _tid, _case in cases.items():
+            if _case.get("status") != "ok":
+                continue
+            _pidx = _case.get("problem_idx")
+            if _pidx is None:
+                _p_parts = _tid.rsplit("_", 2)
+                _pidx = int(_p_parts[-1]) if len(_p_parts) >= 3 else 0
+            problem_groups.setdefault(_pidx, []).append(_case)
+
+        summary_rows_html = ""
+        if problem_groups:
+            def _stat(v, fmt=".1f", avg_fmt=None):
+                if not v:
+                    return ("–", "–", "–")
+                af = fmt if avg_fmt is None else avg_fmt
+                return (f"{min(v):{fmt}}", f"{sum(v)/len(v):{af}}", f"{max(v):{fmt}}")
+
+            summary_data = []
+            for pidx, g in problem_groups.items():
+                runs = len(g)
+                n_ok = sum(1 for c in g if c.get("correct", False))
+                toks = [c["tokens"] for c in g if c.get("tokens") is not None]
+                tps = [c["tps_gen"] for c in g if c.get("tps_gen") is not None]
+                tg = [c["t_gen_ms"] / 1000 for c in g if c.get("t_gen_ms") is not None]
+                summary_data.append((
+                    pidx, runs, n_ok,
+                    _stat(toks, "d", ".0f"),
+                    _stat(tps),
+                    _stat(tg),
+                ))
+
+            summary_data.sort(key=lambda r: r[0])  # sort by problem index ascending
+
+            summary_rows_html = "\n".join(
+                f"""<tr class="summary-row">
+                    <td>{p:03d}</td>
+                    <td>{r}</td>
+                    <td>{n}/{r}</td>
+                    <td>{tk[0]}</td><td>{tk[1]}</td><td>{tk[2]}</td>
+                    <td>{tp[0]}</td><td>{tp[1]}</td><td>{tp[2]}</td>
+                    <td>{tg[0]}</td><td>{tg[1]}</td><td>{tg[2]}</td>
+                </tr>"""
+                for p, r, n, tk, tp, tg in summary_data
+            )
+
        html_content = f"""<!DOCTYPE html>
 <html>
 <head>
@@ -412,10 +471,10 @@ class EvalState:
 <title>{self.dataset_type.upper()} Eval</title>
 <style>
        body {{ font-family: system-ui, sans-serif; margin: 0; padding: 16px; background: #fff; color: #222; }}
-        .bar {{ padding: 8px 0; font-size: 14px; color: #555; }}
-        .bar span {{ margin-right: 20px; }}
-        .bar b {{ color: #222; }}
-        table {{ width: 100%; border-collapse: collapse; font-size: 13px; }}
+        .bar {{ padding: 8px 0; font-size: 13px; color: #555; font-family: 'SF Mono', 'Menlo', 'Consolas', monospace; display: grid; grid-template-columns: auto 1fr auto 1fr; gap: 2px 12px; align-items: baseline; }}
+        .bar .label {{ color: #888; }}
+        .bar .value {{ color: #222; }}
+        table {{ width: 100%; border-collapse: collapse; font-size: 13px; font-family: 'SF Mono', 'Menlo', 'Consolas', monospace; }}
        th {{ text-align: left; padding: 6px 8px; border-bottom: 2px solid #ccc; font-weight: 600; }}
        td {{ padding: 4px 8px; border-bottom: 1px solid #eee; vertical-align: top; }}
        .task-row {{ cursor: pointer; }}
@@ -429,37 +488,88 @@ class EvalState:
        .details-content {{ padding: 8px 16px; background: #f6f8fa; font-size: 12px; }}
        .details-content b {{ color: #555; }}
        .details-content pre {{ background: #fff; border: 1px solid #e1e4e8; padding: 8px; overflow-x: auto; white-space: pre-wrap; word-wrap: break-word; margin: 4px 0 8px; }}
+        .summary-table {{ margin-bottom: 16px; font-size: 13px; width: 100%; }}
+        .summary-row {{ background: #fafbfc; }}
+        .summary-row:hover {{ background: #f5f5f5; }}
+        .summary-table th {{ text-align: right; font-weight: 600; }}
+        .summary-table th:first-child {{ text-align: left; }}
+        .summary-table th[colspan] {{ text-align: center; }}
+        .summary-table td {{ text-align: right; }}
+        .summary-table td:first-child {{ text-align: left; }}
+        .tabs {{ display: flex; border-bottom: 2px solid #ddd; margin: 12px 0 0; }}
+        .tab-btn {{ padding: 6px 16px; border: none; background: none; font-size: 13px; cursor: pointer; color: #555; border-bottom: 2px solid transparent; margin-bottom: -2px; font-weight: 500; }}
+        .tab-btn:hover {{ color: #222; }}
+        .tab-btn.active {{ color: #222; border-bottom-color: #222; font-weight: 600; }}
+        .tab-content {{ display: none; }}
+        .tab-content.active {{ display: block; }}
 </style>
 </head>
 <body>
    <div class="bar">
-        <span><b>{self.dataset_type.upper()}</b></span>
-        <span>Model: {self.model_name or 'N/A'}</span>
-        <span>Accuracy: <b>{accuracy:.1f}%</b> [{ci_lower*100:.1f}%, {ci_upper*100:.1f}%]</span>
-        <span>Correct: <span class="correct">{n_correct}</span> / {len(completed)}</span>
-        <span>Pending: {n_pending}</span>
-        <span>Time: {self.total_time:.1f}s</span>
-        <span>Sampling: {sampling_str}</span>
+        <div class="label">Dataset</div><div class="value"><b>{self.dataset_type.upper()}</b></div>
+        <div class="label">Model</div><div class="value"><b>{self.model_name or 'N/A'}</b></div>
+        <div class="label">Accuracy</div><div class="value"><b>{accuracy:.1f}%</b> [{ci_lower*100:.1f}%, {ci_upper*100:.1f}%]</div>
+        <div class="label">Correct</div><div class="value"><span class="correct">{n_correct}</span> / {len(completed)}</div>
+        <div class="label">Pending</div><div class="value">{n_pending}</div>
+        <div class="label">Time</div><div class="value">{self.total_time:.1f}s</div>
+        <div class="label">Sampling</div><div class="value">{sampling_str}</div>
+    </div>
+    <div class="tabs">
+        <button class="tab-btn active" data-tab="detailed" onclick="switchTab(this)">Detailed</button>
+        <button class="tab-btn" data-tab="summary" onclick="switchTab(this)">Summary</button>
+    </div>
+    <div id="tab-detailed" class="tab-content active">
+        <table>
+            <thead>
+                <tr>
+                    <th>ID</th>
+                    <th></th>
+                    <th>Gold</th>
+                    <th>Answer</th>
+                    <th>Tokens</th>
+                    <th>T/s</th>
+                    <th>Gen s</th>
+                    <th>Server</th>
+                </tr>
+            </thead>
+            <tbody>
+                {rows_html}
+            </tbody>
+        </table>
+    </div>
+    <div id="tab-summary" class="tab-content">
+        <table class="summary-table">
+            <thead>
+                <tr>
+                    <th>Problem</th>
+                    <th>Runs</th>
+                    <th>Correct</th>
+                    <th colspan="3">Tokens</th>
+                    <th colspan="3">T/s</th>
+                    <th colspan="3">Gen s</th>
+                </tr>
+                <tr>
+                    <th></th>
+                    <th></th>
+                    <th></th>
+                    <th>min</th><th>avg</th><th>max</th>
+                    <th>min</th><th>avg</th><th>max</th>
+                    <th>min</th><th>avg</th><th>max</th>
+                </tr>
+            </thead>
+            <tbody>
+                {summary_rows_html}
+            </tbody>
+        </table>
    </div>
-    <table>
-        <thead>
-            <tr>
-                <th>ID</th>
-                <th></th>
-                <th>Gold</th>
-                <th>Answer</th>
-                <th>Tokens</th>
-                <th>T/s</th>
-                <th>Gen s</th>
-                <th>Server</th>
-            </tr>
-        </thead>
-        <tbody>
-            {rows_html}
-        </tbody>
-    </table>
    <script>
        function toggleDetails(id) {{ document.getElementById('details-'+id).classList.toggle('open'); }}
+        function switchTab(btn) {{
+            document.querySelectorAll('.tab-btn').forEach(b => b.classList.remove('active'));
+            document.querySelectorAll('.tab-content').forEach(c => c.classList.remove('active'));
+            btn.classList.add('active');
+            document.getElementById('tab-'+btn.dataset.tab).classList.add('active');
+        }}
    </script>
 </body>
 </html>"""
@@ -1062,12 +1172,19 @@ class Processor:
    ) -> TaskState:
        question_text, prompt, expected = eval_state.get_case(i)

+        # Extract chunk_idx from task_id: "{dataset_type}_{chunk_idx:03d}_{index:03d}"
+        _parts = task_id.rsplit("_", 2)
+        chunk_idx = int(_parts[-2]) if len(_parts) >= 3 else 0
+        problem_idx = i
+
        task_state = TaskState(
            task_id=task_id,
            prompt=prompt,
            expected=expected,
            question_text=question_text,
-            server_name=server_config.name
+            server_name=server_config.name,
+            chunk_idx=chunk_idx,
+            problem_idx=problem_idx,
        )

        try:
@@ -1085,7 +1202,8 @@ class Processor:
                eval_state.add_result(
                    task_id, prompt, expected, result, None,
                    {"finish_reason": finish_reason}, False, task_state.status,
-                    tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name
+                    tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name,
+                    chunk_idx, problem_idx,
                )
                eval_state.dump()
                return task_state
@@ -1108,7 +1226,8 @@ class Processor:
            eval_state.add_result(
                task_id, prompt, expected, result, answer,
                grader_log, is_correct, "ok",
-                tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name
+                tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name,
+                chunk_idx, problem_idx,
            )

            eval_state.dump()
@@ -65,34 +65,70 @@ def normalize_number(s: str) -> Optional[int]:
    return int(match.group(0))

 class AimeDataset:
-    def __init__(self, split: str = "train"):
+    def __init__(self, split: str = "train", dataset_type: str = "aime"):
        self.split = split
+        self.dataset_type = dataset_type
        self.questions: List[Dict] = []
        self._load_dataset()

-    def _load_dataset(self):
-        print(f"Loading AIME dataset (split: {self.split})...")
+    def _get_question_text(self, question: Dict) -> str:
+        """Get question text, handling different dataset field names."""
+        return question.get("problem", question.get("question", ""))

-        cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "AI-MO___aimo-validation-aime" / "default" / "0.0.0"
-        if cache_path.exists():
-            print(f"Using cached dataset from {cache_path}")
-            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split, cache_dir=str(cache_path))
+    def _load_dataset(self):
+        if self.dataset_type == "aime":
+            print(f"Loading AIME dataset (split: {self.split})...")
+            cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "AI-MO___aimo-validation-aime" / "default" / "0.0.0"
+            if cache_path.exists():
+                print(f"Using cached dataset from {cache_path}")
+                ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split, cache_dir=str(cache_path))
+            else:
+                ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split)
+        elif self.dataset_type == "aime2025":
+            print(f"Loading AIME2025 dataset...")
+            ds_list = []
+            for config_name in ["AIME2025-I", "AIME2025-II"]:
+                cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "opencompass___AIME2025" / "default" / "0.0.0"
+                if cache_path.exists():
+                    print(f"Using cached dataset from {cache_path}")
+                    ds = datasets.load_dataset("opencompass/AIME2025", config_name, split="test", cache_dir=str(cache_path))
+                else:
+                    ds = datasets.load_dataset("opencompass/AIME2025", config_name, split="test")
+                ds_list.extend(ds)
+            ds = ds_list
        else:
-            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split)
+            raise ValueError(f"Unknown dataset type: {self.dataset_type}")

        self.questions = list(ds)
-        print(f"AIME dataset loaded: {len(self.questions)} questions")
+        print(f"{self.dataset_type} dataset loaded: {len(self.questions)} questions")

    def find_question(self, request_text: str) -> Optional[Dict]:
+        # Strip common template prefixes to get the actual question text
+        # Templates include things like "Solve the following math problem step by step..."
+        # The actual question usually follows a blank line or after the template instruction
+        cleaned = request_text
+        # Split on double newline and take the part that looks like the problem
+        parts = cleaned.split('\n\n')
+        if len(parts) > 1:
+            # Find the part that's longest (likely the actual problem text)
+            problem_parts = [p for p in parts if len(p.strip()) > 100]
+            if problem_parts:
+                cleaned = max(problem_parts, key=lambda x: len(x))
+
        best_match = None
        best_distance = -1
        best_index = -1

        for i, question in enumerate(self.questions):
-            question_text = question["problem"]
-            request_lower = request_text.lower()
+            question_text = self._get_question_text(question)
+            request_lower = cleaned.lower()
            question_lower = question_text.lower()

+            # Check if question text is contained in the cleaned request
+            if question_lower in request_lower or request_lower in question_lower:
+                debug_log(f"DEBUG: Found substring match at index {i}")
+                return question
+
            # Exact match
            if question_lower == request_lower:
                debug_log(f"DEBUG: Found exact match at index {i}")
@@ -118,7 +154,7 @@ class AimeDataset:
            debug_log(f"DEBUG: Found best partial match at index {best_index} with distance {best_distance:.3f}")
            return best_match

-        debug_log(f"DEBUG: No matching question found for: {request_text[:100]}...")
+        debug_log(f"DEBUG: No matching question found for cleaned: {cleaned[:100]}...")
        return None

    def get_answer(self, question: Dict) -> str:
@@ -134,15 +170,16 @@ class Simulator:
        port: int = 8033,
        host: str = "localhost",
        success_rate: float = 0.8,
-        dataset_split: str = "train"
+        dataset_split: str = "train",
+        dataset_type: str = "aime"
    ):
        self.port = port
        self.host = host
        self.success_rate = success_rate
-        self.dataset = AimeDataset(dataset_split)
+        self.dataset = AimeDataset(dataset_split, dataset_type)
        self.eval_state = EvalState(
-            id="aime-2025",
-            tasks=["aime"],
+            id=dataset_type,
+            tasks=[dataset_type],
            task_states={},
            sampling_config={"temperature": 0, "max_tokens": 2048}
        )
@@ -159,6 +196,10 @@ class Simulator:
        else:
            response_text = self._generate_wrong_answer(question)

+        comp_tokens = random.randint(10000, 60000)
+        tps_gen = random.uniform(90.0, 110.0)
+        t_gen_ms = comp_tokens / tps_gen * 1000
+
        return {
            "id": f"chatcmpl-{int(time.time())}",
            "object": "chat.completion",
@@ -176,8 +217,12 @@ class Simulator:
            ],
            "usage": {
                "prompt_tokens": 100,
-                "completion_tokens": 50,
-                "total_tokens": 150
+                "completion_tokens": comp_tokens,
+                "total_tokens": 100 + comp_tokens
+            },
+            "timings": {
+                "predicted_ms": t_gen_ms,
+                "predicted_per_second": tps_gen
            }
        }

@@ -218,6 +263,12 @@ class Simulator:
        return response

 class RequestHandler(BaseHTTPRequestHandler):
+    def do_GET(self):
+        if self.path == "/v1/models":
+            self._send_json({"data": [{"id": "llama", "object": "model"}]}, 200)
+            return
+        self._send_json({"error": "Not found"}, 404)
+
    def do_POST(self):
        if self.path != "/v1/chat/completions":
            self._send_json({"error": "Not found"}, 404)
@@ -280,6 +331,13 @@ def main():
        default=0.8,
        help="Success rate 0-1 (default: 0.8)"
    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="aime",
+        choices=["aime", "aime2025"],
+        help="Dataset type (default: aime)"
+    )
    parser.add_argument(
        "--dataset-split",
        type=str,
@@ -294,7 +352,8 @@ def main():
        port=args.port,
        host=args.host,
        success_rate=args.success_rate,
-        dataset_split=args.dataset_split
+        dataset_split=args.dataset_split,
+        dataset_type=args.dataset
    )

    server = HTTPServer((args.host, args.port), RequestHandler)
@@ -304,7 +363,7 @@ def main():
    print("\n=== llama-server-simulator ===")
    print(f"Server running on http://{args.host}:{args.port}")
    print(f"Success rate: {args.success_rate}")
-    print(f"AIME dataset loaded: {len(simulator.dataset.questions)} questions")
+    print(f"{args.dataset} dataset loaded: {len(simulator.dataset.questions)} questions")
    print("\nPress Ctrl+C to stop\n")

    try: