feat: better benchmarking

Cargo.lock

@@ -5736,6 +5736,7 @@ dependencies = [
 name = "xtask"
 version = "0.1.0"
 dependencies = [
+ "blake3",
  "libc",
  "serde_json",
 ]

apps/mvp-node/tinygrad_worker.py

@@ -751,6 +751,7 @@ def decode_greedy_device_resident(
     request_id: int | None,
     model_id: str | None,
     progress_every: int,
+    decode_started_at: float,
 ) -> list[int]:
     if max_tokens <= 0:
         return []
@@ -797,6 +798,7 @@ def decode_greedy_device_resident(
                 model_id=model_id,
                 token_index=1,
                 prompt_tokens=len(prompt_tokens),
+                first_token_elapsed_ms=int((time.monotonic() - decode_started_at) * 1000),
             )
         elif progress_every > 0 and tokens_generated % progress_every == 0:
             control(
@@ -836,6 +838,7 @@ def infer_prompt(cmd: dict[str, Any]) -> None:
         prompt_chars=len(prompt),
         max_tokens=max_tokens,
     )
+    encode_started = time.monotonic()
     model_prompt, prompt_template = model_prompt_text(prompt)
     control(
         type="PromptEncodeStarted",
@@ -852,8 +855,10 @@ def infer_prompt(cmd: dict[str, Any]) -> None:
         model_prompt_bytes=len(model_prompt.encode("utf-8")),
         prompt_template=prompt_template,
         prompt_tokens=len(prompt_tokens),
+        elapsed_ms=int((time.monotonic() - encode_started) * 1000),
     )
     progress_every = int(os.environ.get("MVP_TOKEN_PROGRESS_EVERY", "16") or "16")
+    decode_started = time.monotonic()
     control(
         type="DecodeStarted",
         request_id=request_id,
@@ -874,6 +879,7 @@ def infer_prompt(cmd: dict[str, Any]) -> None:
             request_id=request_id,
             model_id=loaded.get("model_id"),
             progress_every=progress_every,
+            decode_started_at=decode_started,
         )
     finally:
         stop_cpu_line_sampler(cpu_sampler)
@@ -883,7 +889,9 @@ def infer_prompt(cmd: dict[str, Any]) -> None:
         model_id=loaded.get("model_id"),
         prompt_tokens=len(prompt_tokens),
         tokens_generated=len(generated),
+        elapsed_ms=int((time.monotonic() - decode_started) * 1000),
     )
+    text_decode_started = time.monotonic()
     control(type="TextDecodeStarted", request_id=request_id, model_id=loaded.get("model_id"), tokens_generated=len(generated))
     raw_text = tokenizer.decode(generated) if generated else ""
     text = strip_chat_stop_markers(raw_text)
@@ -893,6 +901,7 @@ def infer_prompt(cmd: dict[str, Any]) -> None:
         model_id=loaded.get("model_id"),
         tokens_generated=len(generated),
         text_bytes=len(text.encode("utf-8")),
+        elapsed_ms=int((time.monotonic() - text_decode_started) * 1000),
     )
     control(
         type="PromptCompleted",
@@ -930,7 +939,11 @@ def install_ring(cmd: dict[str, Any]) -> None:
         type="RingInstalled",
         ring_id=ring_id,
         edge_id=rings[ring_id]["edge_id"],
+        port=rings[ring_id]["port"],
         direction=rings[ring_id]["direction"],
+        data_capacity=rings[ring_id]["data_capacity"],
+        max_extent=rings[ring_id]["max_extent"],
+        alignment=rings[ring_id]["alignment"],
     )
 
 
@@ -1001,6 +1014,7 @@ def materialize_object(payload: bytes, sequence: int, flags: int) -> dict[str, A
         return {
             "kind": "tokens",
             "tokens": tokens,
+            "token_count": token_count,
             "tensor": TensorCls([tokens], dtype="int32").realize(),
             "start_pos": object_start_pos(sequence, token_count, flags),
         }
@@ -1011,6 +1025,7 @@ def materialize_object(payload: bytes, sequence: int, flags: int) -> dict[str, A
         return {
             "kind": "tokens",
             "tokens": tokens,
+            "token_count": token_count,
             "tensor": TensorCls([tokens], dtype="int32").realize(),
             "start_pos": object_start_pos(sequence, token_count, flags),
         }
@@ -1026,18 +1041,20 @@ def materialize_object(payload: bytes, sequence: int, flags: int) -> dict[str, A
     array = np.frombuffer(payload, dtype=np.float16).copy().reshape(1, token_count, hidden_dim)
     return {
         "kind": "activation",
+        "token_count": token_count,
+        "hidden_dim": hidden_dim,
         "tensor": TensorCls(array).realize(),
         "start_pos": object_start_pos(sequence, token_count, flags),
     }
 
 
-
 def ring_readable(cmd: dict[str, Any]) -> None:
     global next_handle
     ring_id = int(cmd["ring_id"])
     ring = rings[ring_id]
     if ring["direction"] != "ingress":
         fatal("WrongRingDirection", ring_id=ring_id, direction=ring["direction"])
+    started = time.monotonic()
     object_id, sequence, extent, flags, payload = parse_record(ring)
     handle = next_handle
     next_handle += 1
@@ -1047,7 +1064,6 @@ def ring_readable(cmd: dict[str, Any]) -> None:
         sequence=sequence,
         extent=extent,
         flags=flags,
-        payload=payload,
     )
     device_objects[handle] = materialized
     control(
@@ -1059,6 +1075,11 @@ def ring_readable(cmd: dict[str, Any]) -> None:
         extent=extent,
         handle_generation=WORKER_GENERATION,
         handle_id=handle,
+        kind=materialized.get("kind"),
+        token_count=materialized.get("token_count"),
+        hidden_dim=materialized.get("hidden_dim"),
+        start_pos=materialized.get("start_pos"),
+        elapsed_ms=int((time.monotonic() - started) * 1000),
     )
 
 
@@ -1092,6 +1113,7 @@ def write_record(ring: dict[str, Any], object_id: int, sequence: int, payload: b
 def execute_step(cmd: dict[str, Any]) -> None:
     if not role:
         fatal("RoleNotConfigured")
+    step_started = time.monotonic()
     handle = int(cmd["input_handle_id"])
     obj = device_objects.get(handle)
     if obj is None:
@@ -1103,35 +1125,41 @@ def execute_step(cmd: dict[str, Any]) -> None:
     if ring["direction"] != "egress":
         fatal("WrongRingDirection", ring_id=output_ring_id, direction=ring["direction"])
     final_stage = bool(cmd.get("final_stage"))
+    input_kind = obj.get("kind")
+    input_extent = int(obj.get("extent", 0))
     execution_backend = "pipeline_stage"
     if not isinstance(model, PipelineStageTinygradModel):
         execution_backend = "full_transformer"
         if not final_stage:
             fatal("FullTransformerNonFinalStageUnsupported", step_id=int(cmd["step_id"]))
-        if obj.get("kind") != "tokens":
+        if input_kind != "tokens":
-            fatal("FullTransformerInputUnsupported", step_id=int(cmd["step_id"]), kind=obj.get("kind"))
+            fatal("FullTransformerInputUnsupported", step_id=int(cmd["step_id"]), kind=input_kind)
         if int(obj.get("flags", 0)) & FLAG_BEGIN_SEQUENCE and hasattr(model, "forward_jit"):
             model.forward_jit.reset()
         token_array = model(obj["tensor"], int(obj.get("start_pos", 0))).realize().numpy().reshape(-1)
         token = int(token_array[0])
         payload = struct.pack("<I", token)
+        output_kind = "token"
         flags = 1 if token == int(loaded.get("eos_token_id", 0)) else 0
     else:
         if final_stage != bool(getattr(model, "final_stage", False)):
             fatal("FinalStageMismatch", command_final_stage=final_stage, model_final_stage=bool(getattr(model, "final_stage", False)))
-        input_tensor = model.token_hidden(obj["tensor"]) if obj.get("kind") == "tokens" else obj["tensor"]
+        input_tensor = model.token_hidden(obj["tensor"]) if input_kind == "tokens" else obj["tensor"]
         hidden = model.forward_hidden(input_tensor, int(obj.get("start_pos", 0)))
         if final_stage:
             token_array = model.next_token(hidden).realize().numpy().reshape(-1)
             token = int(token_array[0])
             payload = struct.pack("<I", token)
+            output_kind = "token"
             flags = 1 if token == int(loaded.get("eos_token_id", 0)) else 0
         else:
             import numpy as np
 
             activation = hidden.realize().numpy().astype(np.float16, copy=False)
             payload = activation.tobytes()
+            output_kind = "activation"
             flags = 0
+    compute_ready = time.monotonic()
     committed = write_record(
         ring,
         int(cmd["output_object_id"]),
@@ -1139,6 +1167,7 @@ def execute_step(cmd: dict[str, Any]) -> None:
         payload,
         flags,
     )
+    write_ready = time.monotonic()
     control(
         type="StepExecuted",
         step_id=int(cmd["step_id"]),
@@ -1147,6 +1176,15 @@ def execute_step(cmd: dict[str, Any]) -> None:
         sequence=int(cmd["output_sequence"]),
         committed_bytes=committed,
         execution_backend=execution_backend,
+        final_stage=final_stage,
+        input_kind=input_kind,
+        input_extent=input_extent,
+        output_kind=output_kind,
+        payload_bytes=len(payload),
+        record_bytes=committed,
+        stage_execution_ms=int((compute_ready - step_started) * 1000),
+        record_write_ms=int((write_ready - compute_ready) * 1000),
+        elapsed_ms=int((write_ready - step_started) * 1000),
     )
 
 
@@ -1157,22 +1195,39 @@ def release_device_object(cmd: dict[str, Any]) -> None:
 
 
 def encode_prompt(cmd: dict[str, Any]) -> None:
+    started = time.monotonic()
     prompt = str(cmd.get("prompt", ""))
     if tokenizer is not None:
         model_prompt, _ = model_prompt_text(prompt)
         tokens = [int(token) for token in tokenizer.encode(model_prompt)]
     else:
+        model_prompt = prompt
         tokens = [int(byte) for byte in prompt.encode("utf-8")] or [0]
-    control(type="PromptEncoded", request_id=cmd.get("request_id"), tokens=tokens)
+    control(
+        type="PromptEncoded",
+        request_id=cmd.get("request_id"),
+        tokens=tokens,
+        prompt_bytes=len(prompt.encode("utf-8")),
+        model_prompt_bytes=len(model_prompt.encode("utf-8")),
+        elapsed_ms=int((time.monotonic() - started) * 1000),
+    )
 
 
 def decode_tokens(cmd: dict[str, Any]) -> None:
+    started = time.monotonic()
     tokens = [int(token) for token in cmd.get("tokens", [])]
     if tokenizer is not None:
         text = strip_chat_stop_markers(tokenizer.decode(tokens))
     else:
         text = "".join(chr(token) if 32 <= token <= 126 else f"<tok:{token}>" for token in tokens)
-    control(type="TokensDecoded", request_id=cmd.get("request_id"), text=text)
+    control(
+        type="TokensDecoded",
+        request_id=cmd.get("request_id"),
+        text=text,
+        tokens=len(tokens),
+        text_bytes=len(text.encode("utf-8")),
+        elapsed_ms=int((time.monotonic() - started) * 1000),
+    )
 
 
 def shutdown_worker(_: dict[str, Any]) -> None:

crates/mvp-system/src/bin/mvp_chat.rs

@@ -61,6 +61,7 @@ const CHAT_LIFECYCLE_CHANNEL: &str = "mvp.chat.lifecycle";
 const CHAT_RUNTIME_CHANNEL: &str = "mvp.chat.runtime";
 const CHAT_PROMPT_CHANNEL: &str = "mvp.chat.prompt";
 const CHAT_COMPONENT_CHANNEL: &str = "mvp.chat.component";
+const CHAT_BENCHMARK_CHANNEL: &str = "mvp.chat.benchmark";
 
 #[derive(Debug)]
 enum PromptInput {
@@ -151,6 +152,7 @@ where
             "gpu_run": config.gpu_run,
         }),
     );
+    progress.emit_benchmark_envelope(&config);
     confirm_vastai_if_needed(&config)?;
     progress.emit(
         CHAT_RUNTIME_CHANNEL,
@@ -331,6 +333,7 @@ struct Config {
     run_id: u64,
     vastai_yes: bool,
     vastai: Option<ResolvedVastAiConfig>,
+    model: ChatModelConfig,
     pipeline_stages: u32,
     max_tokens: u32,
     skip_rebuild: bool,
@@ -379,6 +382,7 @@ impl ChatDatastream {
             CHAT_RUNTIME_CHANNEL,
             CHAT_PROMPT_CHANNEL,
             CHAT_COMPONENT_CHANNEL,
+            CHAT_BENCHMARK_CHANNEL,
         ] {
             out.channel_by_name(name);
         }
@@ -415,6 +419,68 @@ impl ChatDatastream {
         self.flush();
     }
 
+    fn emit_benchmark_envelope(&mut self, config: &Config) {
+        let id = self.channel_by_name(CHAT_BENCHMARK_CHANNEL);
+        let payload = serde_json::to_vec(&json!({
+            "type": "BenchmarkRunEnvelope",
+            "phase": "run_envelope",
+            "status": "ready",
+            "run_id": self.run_id,
+            "benchmark": benchmark_observability::stamp("mvp-chat"),
+            "detail": {
+                "scenario": "mvp-chat",
+                "detail_level": "benchmark_observability_v1",
+                "workload": {
+                    "mode": "stdin_prompt_corpus",
+                    "max_tokens": config.max_tokens,
+                    "prompt_corpus": "external_or_stdin",
+                },
+                "model": {
+                    "id": config.model.id.as_deref(),
+                    "gguf_local_path": config.model.gguf_local_path.as_deref(),
+                    "gguf_repo": config.model.gguf_repo.as_deref(),
+                    "gguf_file": config.model.gguf_file.as_deref(),
+                    "gguf_revision": config.model.gguf_revision.as_deref(),
+                    "tokenizer_local_path": config.model.tokenizer_local_path.as_deref(),
+                    "max_context": config.model.max_context,
+                },
+                "runtime": {
+                    "provider": config.provider.as_str(),
+                    "pipeline_stages": config.pipeline_stages,
+                    "orchestrator_launch_mode": config.orchestrator_launch_mode(),
+                    "gpu_run": config.gpu_run,
+                    "relay_mode": config.relay_mode.as_deref(),
+                    "relay_configured": config.relay_url.is_some(),
+                    "endpoint_addr_mask": config.endpoint_addr_mask.as_str(),
+                },
+                "provider": {
+                    "kind": config.provider.as_str(),
+                    "node_image": &config.node_image,
+                    "image_tag": config.image_tag.as_deref(),
+                    "cached_model": config.cached_model.as_ref().map(|model| model.host_path.to_string_lossy().to_string()),
+                    "vastai": config.vastai.as_ref().map(|vastai| json!({
+                        "image": &vastai.image,
+                        "relay_configured": !vastai.relay_url.is_empty(),
+                        "bootstrap_command_configured": !vastai.bootstrap_command.is_empty(),
+                        "gpu_name": vastai.gpu_name.as_deref(),
+                        "min_gpu_ram_mb": vastai.min_gpu_ram_mb,
+                        "min_down_mbps": vastai.min_down_mbps,
+                        "min_up_mbps": vastai.min_up_mbps,
+                        "max_dph_total": vastai.max_dph_total,
+                        "min_reliability": vastai.min_reliability,
+                        "require_verified": vastai.require_verified,
+                        "disk_gb": vastai.disk_gb,
+                        "has_onstart": vastai.onstart.is_some(),
+                        "has_ssh_identity": vastai.ssh_identity.is_some(),
+                    })),
+                },
+            },
+        }))
+        .expect("serialize mvp-chat benchmark envelope");
+        self.producer.submit_bytes(id, payload);
+        self.flush();
+    }
+
     fn flush(&mut self) {
         let stream = self.stream.clone();
         for frame in self.endpoint.mux().drain() {
@@ -521,6 +587,7 @@ struct ChatTomlConfig {
     observability: ChatObservabilityConfig,
     image: ChatImageConfig,
     vastai: ChatVastAiConfig,
+    model: ChatModelConfig,
     relay: ChatRelayConfig,
 }
 
@@ -568,6 +635,7 @@ struct ChatVastAiConfig {
     min_gpu_ram_mb: Option<u64>,
     min_down_mbps: Option<f64>,
     min_up_mbps: Option<f64>,
+    max_dph_total: Option<f64>,
     min_reliability: Option<f64>,
     require_verified: Option<bool>,
     disk_gb: Option<u32>,
@@ -575,6 +643,18 @@ struct ChatVastAiConfig {
     ssh_identity: Option<String>,
 }
 
+#[derive(Clone, Debug, Default, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+struct ChatModelConfig {
+    id: Option<String>,
+    gguf_local_path: Option<String>,
+    gguf_repo: Option<String>,
+    gguf_file: Option<String>,
+    gguf_revision: Option<String>,
+    tokenizer_local_path: Option<String>,
+    max_context: Option<u32>,
+}
+
 #[derive(Clone, Debug)]
 struct LoadedChatTomlConfig {
     overlay: ChatTomlConfig,
@@ -687,6 +767,7 @@ impl Config {
             vastai_yes: args.vastai_yes,
             pipeline_stages,
             max_tokens,
+            model: toml.model,
             vastai,
             skip_rebuild: args.skip_rebuild,
             gpu_run,
@@ -714,6 +795,27 @@ impl Config {
             self.pipeline_stages.to_string(),
             "--no-dashboard".to_owned(),
         ];
+        if let Some(model_id) = &self.model.id {
+            args.extend(["--model-id".to_owned(), model_id.clone()]);
+        }
+        if let Some(path) = &self.model.gguf_local_path {
+            args.extend(["--gguf-local-path".to_owned(), path.clone()]);
+        }
+        if let Some(repo) = &self.model.gguf_repo {
+            args.extend(["--gguf-repo".to_owned(), repo.clone()]);
+        }
+        if let Some(file) = &self.model.gguf_file {
+            args.extend(["--gguf-file".to_owned(), file.clone()]);
+        }
+        if let Some(revision) = &self.model.gguf_revision {
+            args.extend(["--gguf-revision".to_owned(), revision.clone()]);
+        }
+        if let Some(path) = &self.model.tokenizer_local_path {
+            args.extend(["--tokenizer-local-path".to_owned(), path.clone()]);
+        }
+        if let Some(max_context) = self.model.max_context {
+            args.extend(["--max-context".to_owned(), max_context.to_string()]);
+        }
         if self.provider == ProviderKind::Process {
             args.extend([
                 "--worker-bin".to_owned(),
@@ -773,6 +875,12 @@ impl Config {
             if let Some(min_up_mbps) = vastai.min_up_mbps {
                 args.extend(["--vastai-min-up-mbps".to_owned(), min_up_mbps.to_string()]);
             }
+            if let Some(max_dph_total) = vastai.max_dph_total {
+                args.extend([
+                    "--vastai-max-dph-total".to_owned(),
+                    max_dph_total.to_string(),
+                ]);
+            }
             if let Some(min_reliability) = vastai.min_reliability {
                 args.extend([
                     "--vastai-min-reliability".to_owned(),
@@ -919,6 +1027,7 @@ fn resolve_vastai_config(
         min_gpu_ram_mb: file.min_gpu_ram_mb,
         min_down_mbps: file.min_down_mbps,
         min_up_mbps: file.min_up_mbps,
+        max_dph_total: file.max_dph_total,
         min_reliability: file.min_reliability,
         require_verified: file.require_verified,
         onstart: first_non_empty([file.onstart.clone()]),
@@ -1362,6 +1471,23 @@ fn prepare_runtime_with_progress(
     }
 
     if config.skip_rebuild {
+        if config.provider == ProviderKind::VastAi {
+            emit_chat_progress(
+                &mut progress,
+                CHAT_RUNTIME_CHANNEL,
+                "ensure_worker_binary",
+                "skipped",
+                json!({"mode": binary_mode, "reason": "vastai_remote_image"}),
+            );
+            emit_chat_progress(
+                &mut progress,
+                CHAT_RUNTIME_CHANNEL,
+                "prepare_node_image",
+                "skipped",
+                json!({"provider": config.provider.as_str(), "reason": "skip_rebuild"}),
+            );
+            return Ok(config.node_image.clone());
+        }
         emit_chat_progress(
             &mut progress,
             CHAT_RUNTIME_CHANNEL,
@@ -1601,6 +1727,10 @@ fn emit_chat_progress(
     }
 }
 
+fn prompt_hash_hex(prompt: &str) -> String {
+    blake3::hash(prompt.as_bytes()).to_hex().to_string()
+}
+
 fn run_chat_session_with_output_and_progress<R, W, O>(
     writer: &mut W,
     mut reader: R,
@@ -1616,6 +1746,7 @@ where
 {
     let mut progress = progress;
     let mut next_request_id = 1_u64;
+    let mut next_prompt_index = 1_u64;
 
     loop {
         if STOP_REQUESTED.load(Ordering::SeqCst) {
@@ -1633,7 +1764,7 @@ where
             CHAT_PROMPT_CHANNEL,
             "waiting_for_prompt",
             "started",
-            json!({"next_request_id": next_request_id}),
+            json!({"next_request_id": next_request_id, "next_prompt_index": next_prompt_index}),
         );
         write!(output, "prompt:> ").map_err(|e| format!("write prompt: {e}"))?;
         output.flush().map_err(|e| format!("flush prompt: {e}"))?;
@@ -1666,12 +1797,15 @@ where
 
         let request_id = next_request_id;
         next_request_id = next_request_id.wrapping_add(1).max(1);
+        let prompt_index = next_prompt_index;
+        next_prompt_index = next_prompt_index.wrapping_add(1).max(1);
+        let prompt_hash = prompt_hash_hex(&prompt);
         emit_chat_progress(
             &mut progress,
             CHAT_PROMPT_CHANNEL,
             "prompt_submitted",
             "ready",
-            json!({"request_id": request_id, "prompt_bytes": prompt.len(), "max_tokens": max_tokens}),
+            json!({"request_id": request_id, "prompt_index": prompt_index, "prompt_hash": &prompt_hash, "prompt_bytes": prompt.len(), "max_tokens": max_tokens}),
         );
         write_json_line(
             writer,
@@ -1687,7 +1821,7 @@ where
             CHAT_PROMPT_CHANNEL,
             "decoding",
             "started",
-            json!({"request_id": request_id}),
+            json!({"request_id": request_id, "prompt_index": prompt_index, "prompt_hash": &prompt_hash}),
         );
         let mut response_started = false;
 
@@ -1768,7 +1902,7 @@ where
                         CHAT_PROMPT_CHANNEL,
                         "response_text",
                         "observed",
-                        json!({"request_id": request_id, "text_bytes": text.len()}),
+                        json!({"request_id": request_id, "prompt_index": prompt_index, "prompt_hash": &prompt_hash, "text_bytes": text.len()}),
                     );
                 }
                 PromptEvent::Done {
@@ -1790,6 +1924,8 @@ where
                         "ready",
                         json!({
                             "request_id": request_id,
+                            "prompt_index": prompt_index,
+                            "prompt_hash": &prompt_hash,
                             "response_started": response_started,
                             "tokens_generated": tokens_generated,
                             "elapsed_ms": elapsed_ms,
@@ -1806,7 +1942,7 @@ where
                         CHAT_PROMPT_CHANNEL,
                         "request_faulted",
                         "ready",
-                        json!({"request_id": request_id, "error": error}),
+                        json!({"request_id": request_id, "prompt_index": prompt_index, "prompt_hash": &prompt_hash, "error": error}),
                     );
                     break;
                 }
@@ -2185,6 +2321,7 @@ mod tests {
             run_id: 1,
             vastai_yes: false,
             vastai: None,
+            model: ChatModelConfig::default(),
             pipeline_stages: 1,
             max_tokens: DEFAULT_MAX_TOKENS,
             skip_rebuild: true,
@@ -2206,6 +2343,7 @@ mod tests {
             min_gpu_ram_mb: None,
             min_down_mbps: None,
             min_up_mbps: None,
+            max_dph_total: None,
             min_reliability: None,
             require_verified: None,
             onstart: None,
@@ -2934,6 +3072,25 @@ relay_url = "https://relay.example"
         assert_eq!(image_ref, "docker.io/acme/node:latest");
     }
 
+    #[test]
+    fn vastai_skip_rebuild_uses_remote_image_without_worker_artifact() {
+        let temp = TempDir::new("vastai-skip-rebuild");
+        let orch_bin = temp.path().join("mvp-orchestrator");
+        let worker_bin = temp.path().join("mvp-worker-node");
+        let mut config = base_config(ProviderKind::VastAi);
+        config.skip_rebuild = true;
+        config.orch_bin = orch_bin.clone();
+        config.worker_bin = worker_bin;
+        config.node_image = "docker.io/acme/node:latest".to_owned();
+
+        assert!(prepare_runtime_with(&config, panic_prepare_node_image).is_err());
+
+        fs::write(&orch_bin, b"orch").expect("write orchestrator artifact");
+        let image_ref = prepare_runtime_with(&config, panic_prepare_node_image)
+            .expect("VastAI skip rebuild reuses remote image");
+        assert_eq!(image_ref, "docker.io/acme/node:latest");
+    }
+
     #[test]
     fn artifact_roots_use_current_directory() {
         let temp = TempDir::new("artifact-root");

crates/mvp-system/src/bin/worker_node.rs

@@ -592,6 +592,14 @@ impl WorkerEdgeRuntime {
                             edge_id: driver_model::EdgeId(edge_id),
                             stream_id: driver_model::StreamId(stream_id),
                         });
+                    emit_node_event(
+                        datastream,
+                        config,
+                        NODE_STAGE_CHANNEL,
+                        "iroh_edge_stream_arrived",
+                        "observed",
+                        json!({"edge_id":edge_id,"stream_id":stream_id}),
+                    );
                     self.drive_edge_workflow(
                         stack,
                         node_actor,
@@ -608,6 +616,15 @@ impl WorkerEdgeRuntime {
                     bytes,
                     ..
                 } => {
+                    let byte_count = bytes.len();
+                    emit_node_event(
+                        datastream,
+                        config,
+                        NODE_STAGE_CHANNEL,
+                        "iroh_edge_bytes_read",
+                        "observed",
+                        json!({"edge_id":edge_id,"stream_id":stream_id,"bytes":byte_count}),
+                    );
                     self.ingest_stream_bytes(
                         edge_id,
                         stream_id,
@@ -792,6 +809,22 @@ impl WorkerEdgeRuntime {
                 .read_arena(lease.layout.data_offset, committed_bytes)
                 .map_err(|e| format!("read egress ring: {e}"))?
         };
+        let record_bytes = record.len();
+        emit_node_event(
+            datastream,
+            config,
+            NODE_STAGE_CHANNEL,
+            "egress_ring_read",
+            "ready",
+            json!({
+                "edge_id":outbound.edge_id,
+                "edge_kind":format!("{:?}", outbound.kind),
+                "ring_id":output_ring_id,
+                "record_bytes":record_bytes,
+                "committed_bytes":committed_bytes,
+                "final_stage":final_stage,
+            }),
+        );
         self.driver_model
             .observe(driver_model::DriverEvent::EgressBytesCommitted {
                 edge_id: driver_model::EdgeId(outbound.edge_id),
@@ -806,6 +839,19 @@ impl WorkerEdgeRuntime {
             .as_ref()
             .ok_or_else(|| "outbound edge sender missing".to_owned())?;
         sender.send(record)?;
+        emit_node_event(
+            datastream,
+            config,
+            NODE_STAGE_CHANNEL,
+            "iroh_edge_bytes_sent",
+            "ready",
+            json!({
+                "edge_id":outbound.edge_id,
+                "edge_kind":format!("{:?}", outbound.kind),
+                "bytes":record_bytes,
+                "record_bytes":record_bytes,
+            }),
+        );
         stack
             .runtime
             .send_to(node_actor, NodeAgentMsg::StepCompleted { step_id })
@@ -867,6 +913,19 @@ impl WorkerEdgeRuntime {
                     .write_arena(lease.layout.data_offset, &record)
                     .map_err(|e| format!("write ingress ring: {e}"))?;
             }
+            emit_node_event(
+                datastream,
+                config,
+                NODE_STAGE_CHANNEL,
+                "ingress_ring_write",
+                "ready",
+                json!({
+                    "edge_id":edge_id,
+                    "edge_kind":format!("{:?}", inbound.kind),
+                    "ring_id":ring_id,
+                    "record_bytes":record.len(),
+                }),
+            );
             let loaded = worker.ring_readable(
                 ring_id,
                 edge_id,

crates/mvp-system/src/config.rs

@@ -106,6 +106,7 @@ pub struct VastAiConfig {
     pub min_gpu_ram_mb: Option<u64>,
     pub min_down_mbps: Option<f64>,
     pub min_up_mbps: Option<f64>,
+    pub max_dph_total: Option<f64>,
     pub min_reliability: Option<f64>,
     pub require_verified: Option<bool>,
     pub poll_interval_secs: Option<u64>,
@@ -130,6 +131,7 @@ pub struct ResolvedVastAiConfig {
     pub min_gpu_ram_mb: Option<u64>,
     pub min_down_mbps: Option<f64>,
     pub min_up_mbps: Option<f64>,
+    pub max_dph_total: Option<f64>,
     pub min_reliability: Option<f64>,
     pub require_verified: Option<bool>,
     pub onstart: Option<String>,
@@ -204,6 +206,9 @@ impl ResolvedVastAiConfig {
         if let Some(min_up_mbps) = self.min_up_mbps {
             policy.min_up_mbps = Some(min_up_mbps);
         }
+        if let Some(max_dph_total) = self.max_dph_total {
+            policy.max_dph_total = Some(max_dph_total);
+        }
         if let Some(min_reliability) = self.min_reliability {
             policy.min_reliability = min_reliability;
         }
@@ -259,6 +264,7 @@ mod tests {
             min_gpu_ram_mb: Some(16_000),
             min_down_mbps: Some(100.0),
             min_up_mbps: Some(25.0),
+            max_dph_total: Some(0.10),
             min_reliability: Some(0.98),
             require_verified: Some(true),
             onstart: None,
@@ -326,6 +332,7 @@ gpu_name = "RTX 4090"
 min_gpu_ram_mb = 24000
 min_down_mbps = 250.5
 min_up_mbps = 50.25
+max_dph_total = 0.10
 min_reliability = 0.99
 require_verified = true
 onstart = "echo preparing"
@@ -402,6 +409,7 @@ poll_interval_secs = 30
         assert_eq!(config.vastai.min_gpu_ram_mb, Some(24_000));
         assert_eq!(config.vastai.min_down_mbps, Some(250.5));
         assert_eq!(config.vastai.min_up_mbps, Some(50.25));
+        assert_eq!(config.vastai.max_dph_total, Some(0.10));
         assert_eq!(config.vastai.min_reliability, Some(0.99));
         assert_eq!(config.vastai.require_verified, Some(true));
         assert_eq!(config.vastai.poll_interval_secs, Some(30));

crates/mvp-system/src/orchestrator_app.rs

@@ -636,6 +636,15 @@ impl VastAiRuntimeConfig {
         if let Some(min_down_mbps) = min_down_mbps {
             provisioning.selection.min_down_mbps = min_down_mbps;
         }
+        let max_dph_total = builder
+            .vastai_max_dph_total_raw
+            .as_ref()
+            .map(|value| ConfigBuilder::parse_value("MVP_VASTAI_MAX_DPH_TOTAL", value))
+            .transpose()?
+            .or(builder.vastai_max_dph_total);
+        if let Some(max_dph_total) = max_dph_total {
+            provisioning.selection.max_dph_total = Some(max_dph_total);
+        }
         let min_up_mbps = builder
             .vastai_min_up_mbps_raw
             .as_ref()
@@ -695,6 +704,7 @@ impl VastAiRuntimeConfig {
             "min_gpu_ram_mb": self.provisioning.selection.min_gpu_ram_mb,
             "min_down_mbps": self.provisioning.selection.min_down_mbps,
             "min_up_mbps": self.provisioning.selection.min_up_mbps,
+            "max_dph_total": self.provisioning.selection.max_dph_total,
             "min_reliability": self.provisioning.selection.min_reliability,
             "require_verified": self.provisioning.selection.require_verified,
             "confirm_lease": self.provisioning.confirm_lease,
@@ -910,6 +920,8 @@ struct ConfigBuilder {
     vastai_min_down_mbps: Option<f64>,
     vastai_min_down_mbps_raw: Option<String>,
     vastai_min_up_mbps: Option<f64>,
+    vastai_max_dph_total: Option<f64>,
+    vastai_max_dph_total_raw: Option<String>,
     vastai_min_up_mbps_raw: Option<String>,
     vastai_min_reliability: Option<f64>,
     vastai_min_reliability_raw: Option<String>,
@@ -965,6 +977,8 @@ impl ConfigBuilder {
             vastai_min_gpu_ram_mb_raw: None,
             vastai_min_down_mbps: None,
             vastai_min_down_mbps_raw: None,
+            vastai_max_dph_total: None,
+            vastai_max_dph_total_raw: None,
             vastai_min_up_mbps: None,
             vastai_min_up_mbps_raw: None,
             vastai_min_reliability: None,
@@ -1083,6 +1097,9 @@ impl ConfigBuilder {
         if let Some(min_down_mbps) = overlay.vastai.min_down_mbps {
             self.vastai_min_down_mbps = Some(min_down_mbps);
         }
+        if let Some(max_dph_total) = overlay.vastai.max_dph_total {
+            self.vastai_max_dph_total = Some(max_dph_total);
+        }
         if let Some(min_up_mbps) = overlay.vastai.min_up_mbps {
             self.vastai_min_up_mbps = Some(min_up_mbps);
         }
@@ -1215,6 +1232,9 @@ impl ConfigBuilder {
         if let Some(min_down_mbps) = env_optional("MVP_VASTAI_MIN_DOWN_MBPS") {
             self.vastai_min_down_mbps_raw = Some(min_down_mbps);
         }
+        if let Some(max_dph_total) = env_optional("MVP_VASTAI_MAX_DPH_TOTAL") {
+            self.vastai_max_dph_total_raw = Some(max_dph_total);
+        }
         if let Some(min_up_mbps) = env_optional("MVP_VASTAI_MIN_UP_MBPS") {
             self.vastai_min_up_mbps_raw = Some(min_up_mbps);
         }
@@ -1339,6 +1359,11 @@ impl ConfigBuilder {
                         Some(parse_next(&mut args, "--vastai-min-down-mbps")?);
                     self.vastai_min_down_mbps_raw = None;
                 }
+                "--vastai-max-dph-total" => {
+                    self.vastai_max_dph_total =
+                        Some(parse_next(&mut args, "--vastai-max-dph-total")?);
+                    self.vastai_max_dph_total_raw = None;
+                }
                 "--vastai-min-up-mbps" => {
                     self.vastai_min_up_mbps = Some(parse_next(&mut args, "--vastai-min-up-mbps")?);
                     self.vastai_min_up_mbps_raw = None;
@@ -1380,9 +1405,9 @@ impl ConfigBuilder {
         if self.pipeline_stages == 0 {
             return Err("--pipeline-stages must be greater than 0".to_owned());
         }
-        if provider == ProviderKind::VastAi && self.pipeline_stages > 1 {
+        if provider == ProviderKind::VastAi && self.pipeline_stages > 2 {
             return Err(
-                "pipeline stages greater than 1 are only supported with provider=docker; provider=vastai does not support pipelined provisioning yet".to_owned(),
+                "provider=vastai currently supports at most 2 pipeline stages for activation-path smoke checks".to_owned(),
             );
         }
         let mut cached_model_host_path = self.cached_model_host_path.clone();
@@ -1648,6 +1673,20 @@ impl Config {
                     ))
                 }
             }
+            GgufSource::HuggingFaceGguf { repo, file, .. }
+                if self.provider == ProviderKind::VastAi
+                    && gguf_source_matches_default_pipeline_cache(&self.gguf_source) =>
+            {
+                let host_path = default_pipeline_cached_model_path();
+                if host_path.is_file() {
+                    Ok(host_path)
+                } else {
+                    Err(format!(
+                        "VastAI pipeline planning requires local GGUF metadata at {}; selected remote source is {repo}/{file}",
+                        host_path.display()
+                    ))
+                }
+            }
             GgufSource::HuggingFaceGguf { repo, file, .. } => Err(format!(
                 "local pipeline requires a locally inspectable GGUF before provisioning; selected source {repo}/{file} is remote, so use --cached-model-host-path"
             )),
@@ -6618,25 +6657,91 @@ kind = "docker"
     }
 
     #[test]
-    fn vastai_rejects_pipeline_stages_count_above_one() {
+    fn vastai_rejects_pipeline_stages_count_above_two() {
         let error = with_clean_env(&[], || {
             match Config::from_layers_with_path_and_args(
                 None,
-                ["--provider", "vastai", "--pipeline-stages", "2"]
+                ["--provider", "vastai", "--pipeline-stages", "3"]
                     .into_iter()
                     .map(str::to_owned),
             ) {
-                Ok(_) => panic!("VastAI cannot provision more than one pipeline stage yet"),
+                Ok(_) => panic!("VastAI smoke runs are capped at two pipeline stages"),
                 Err(error) => error,
             }
         });
 
         assert!(
-            error.contains("provider=vastai does not support pipelined provisioning yet"),
+            error.contains("provider=vastai currently supports at most 2 pipeline stages"),
             "unexpected error: {error}"
         );
     }
 
+    #[test]
+    fn vastai_two_stage_plan_uses_remote_gguf_and_no_mounts() {
+        let config = with_clean_env(&[], || {
+            Config::from_layers_with_path_and_args(
+                None,
+                [
+                    "--provider",
+                    "vastai",
+                    "--pipeline-stages",
+                    "2",
+                    "--model-id",
+                    "smollm2-135m-instruct-q4",
+                    "--gguf-repo",
+                    "QuantFactory/SmolLM2-135M-Instruct-GGUF",
+                    "--gguf-file",
+                    DEFAULT_PIPELINE_CACHED_MODEL_FILE,
+                    "--max-context",
+                    "256",
+                    "--relay-mode",
+                    "default",
+                    "--vastai-bootstrap-command",
+                    "boot",
+                ]
+                .into_iter()
+                .map(str::to_owned),
+            )
+            .expect("VastAI two-stage pipeline config parses")
+        });
+        let plan = config
+            .build_run_plan()
+            .expect("VastAI two-stage run plan uses local metadata only");
+        let coordinator = EndpointAddr::new(iroh::SecretKey::from_bytes(&[41; 32]).public());
+        let orchestrator_actor = ActorAddress([42; 32]);
+
+        let specs = stage_node_specs(&config, Some(&plan), coordinator, orchestrator_actor)
+            .expect("VastAI pipeline stage node specs build");
+
+        assert_eq!(config.provider, ProviderKind::VastAi);
+        assert!(
+            config.cached_model.is_none(),
+            "VastAI must not mount host caches"
+        );
+        assert_eq!(specs.len(), 2);
+        for (expected_stage_index, spec) in specs.iter().enumerate() {
+            let expected_stage_index =
+                u32::try_from(expected_stage_index).expect("fixture stage index fits u32");
+            let expected_node_id = config.node_id + 1 + u64::from(expected_stage_index);
+            assert_eq!(spec.node_id, expected_node_id);
+            assert_eq!(spec.stage_index, Some(expected_stage_index));
+            assert_eq!(env_value(&spec.env, "MVP_NODE_PROVIDER"), Some("vastai"));
+            assert_eq!(env_value(&spec.env, "MVP_PIPELINE_STAGES"), Some("2"));
+            assert_eq!(
+                env_value(&spec.env, "MVP_GGUF_REPO"),
+                Some("QuantFactory/SmolLM2-135M-Instruct-GGUF")
+            );
+            assert_eq!(
+                env_value(&spec.env, "MVP_GGUF_FILE"),
+                Some(DEFAULT_PIPELINE_CACHED_MODEL_FILE)
+            );
+            assert_eq!(env_value(&spec.env, "MVP_GGUF_LOCAL_PATH"), None);
+            assert_eq!(env_value(&spec.env, "MVP_MAX_CONTEXT"), Some("256"));
+            assert_eq!(spec.args, vec!["boot".to_owned()]);
+            assert!(spec.mounts.is_empty(), "VastAI stage specs must not mount");
+        }
+    }
+
     #[test]
     fn pipeline_cached_model_resolution_uses_toml_smollm2_gguf_file_with_cli_stage_count() {
         let toml = TempTomlFile::new(

crates/mvp-system/src/vastai_provisioning.rs

@@ -797,6 +797,24 @@ where
                 instance.contract_id
             ),
         });
+        sink.observe(PluginObservation::ProviderLine {
+            run_id: spec.run_id,
+            node_id: spec.node_id,
+            line: serde_json::json!({
+                "type": "VastAiLeaseReady",
+                "run_id": spec.run_id,
+                "node_id": spec.node_id,
+                "label": &label,
+                "image": &spec.image,
+                "contract_id": instance.contract_id,
+                "offer_id": instance.offer_id,
+                "host_id": instance.host_id,
+                "gpu_name": &instance.gpu_name,
+                "gpu_ram": instance.gpu_ram,
+                "dph_total": instance.dph_total,
+            })
+            .to_string(),
+        });
 
         let endpoint = match self.client.ssh_endpoint(
             instance.contract_id,
@@ -812,6 +830,20 @@ where
                 ));
             }
         };
+        sink.observe(PluginObservation::ProviderLine {
+            run_id: spec.run_id,
+            node_id: spec.node_id,
+            line: serde_json::json!({
+                "type": "VastAiSshEndpointReady",
+                "run_id": spec.run_id,
+                "node_id": spec.node_id,
+                "contract_id": instance.contract_id,
+                "host": &endpoint.host,
+                "port": endpoint.port,
+                "user": &endpoint.user,
+            })
+            .to_string(),
+        });
 
         let bootstrap = match self.bootstrap.start_bootstrap(
             spec.clone(),
@@ -843,14 +875,7 @@ where
         Ok(handle)
     }
 
-    fn complete_bootstrap(&mut self, handle: &PluginNodeHandle) -> Result<(), String> {
+    fn complete_bootstrap(&mut self, _handle: &PluginNodeHandle) -> Result<(), String> {
-        let Some(node) = self.nodes.get_mut(&handle.id) else {
-            return Ok(());
-        };
-        if let Some(mut bootstrap) = node.bootstrap.take() {
-            self.bootstrap
-                .stop_bootstrap(&mut bootstrap, BootstrapStopReason::RuntimeReady);
-        }
         Ok(())
     }
 
@@ -937,6 +962,81 @@ mod tests {
         VastAiProvisioningPlugin::new(NoopLeaseClient, NoopBootstrapLauncher, config)
     }
 
+    #[derive(Clone, Default)]
+    struct ObservationSink {
+        observations: Arc<Mutex<Vec<PluginObservation>>>,
+    }
+
+    impl crate::provisioning::PluginObservationSink for ObservationSink {
+        fn observe(&self, observation: PluginObservation) {
+            self.observations.lock().push(observation);
+        }
+    }
+
+    #[derive(Clone)]
+    struct RecordingLeaseClient {
+        destroyed_contracts: Arc<Mutex<Vec<u64>>>,
+    }
+
+    impl VastAiLeaseClient for RecordingLeaseClient {
+        fn provision_one(
+            &mut self,
+            _request: ProvisionRequest,
+        ) -> Result<ProvisionedInstance, String> {
+            Ok(ProvisionedInstance {
+                index: 0,
+                contract_id: 42,
+                offer_id: 7,
+                host_id: Some(99),
+                gpu_name: "RTX 4060".to_owned(),
+                gpu_ram: Some(8_192.0),
+                dph_total: 0.064,
+            })
+        }
+
+        fn ssh_endpoint(
+            &mut self,
+            _contract_id: u64,
+            _label: &str,
+            _lifecycle: &LifecyclePolicy,
+            ssh_user: &str,
+        ) -> Result<VastAiSshEndpoint, String> {
+            Ok(VastAiSshEndpoint {
+                host: "ssh5.vast.ai".to_owned(),
+                port: 22_017,
+                user: ssh_user.to_owned(),
+            })
+        }
+
+        fn destroy_contract(&mut self, contract_id: u64) -> Result<(), String> {
+            self.destroyed_contracts.lock().push(contract_id);
+            Ok(())
+        }
+    }
+
+    #[derive(Clone)]
+    struct RecordingBootstrapLauncher {
+        stop_reasons: Arc<Mutex<Vec<BootstrapStopReason>>>,
+    }
+
+    impl VastAiBootstrapLauncher for RecordingBootstrapLauncher {
+        type Handle = u64;
+
+        fn start_bootstrap(
+            &mut self,
+            spec: NodeProvisionSpec,
+            _endpoint: VastAiSshEndpoint,
+            _sink: PluginSink,
+            _producer: Option<DatastreamProducer>,
+        ) -> Result<Self::Handle, String> {
+            Ok(spec.node_id)
+        }
+
+        fn stop_bootstrap(&mut self, _handle: &mut Self::Handle, reason: BootstrapStopReason) {
+            self.stop_reasons.lock().push(reason);
+        }
+    }
+
     #[test]
     fn vastai_provisioning_build_request_keeps_bootstrap_args_out_of_onstart() {
         let plugin = plugin_with_onstart(None);
@@ -957,6 +1057,38 @@ mod tests {
         assert_eq!(request.onstart.as_deref(), Some("echo explicit setup"));
     }
 
+    #[test]
+    fn vastai_complete_bootstrap_keeps_log_tail_until_node_stop() {
+        let destroyed_contracts = Arc::new(Mutex::new(Vec::new()));
+        let stop_reasons = Arc::new(Mutex::new(Vec::new()));
+        let sink = PluginSink::new(Arc::new(ObservationSink::default()));
+        let mut plugin = VastAiProvisioningPlugin::new(
+            RecordingLeaseClient {
+                destroyed_contracts: destroyed_contracts.clone(),
+            },
+            RecordingBootstrapLauncher {
+                stop_reasons: stop_reasons.clone(),
+            },
+            VastAiProvisioningConfig::default(),
+        );
+
+        let handle = plugin
+            .start_node(node_spec_with_bootstrap_args(), sink)
+            .expect("VastAI node starts");
+
+        plugin
+            .complete_bootstrap(&handle)
+            .expect("runtime-ready bootstrap completion succeeds");
+        assert!(
+            stop_reasons.lock().is_empty(),
+            "VastAI bootstrap SSH tail must remain alive for post-ready worker logs"
+        );
+
+        plugin.stop_node(&handle).expect("VastAI node stops");
+        assert_eq!(*stop_reasons.lock(), vec![BootstrapStopReason::NodeStop]);
+        assert_eq!(*destroyed_contracts.lock(), vec![42]);
+    }
+
     #[test]
     fn vastai_provisioning_next_ssh_backoff_doubles_until_thirty_second_cap() {
         for (current, expected) in [

src/actor.rs

@@ -681,7 +681,6 @@ impl<'a> Ctx<'a> {
         self.inner.extension()
     }
 
-
     /// Return system-level information (worker count, actor count, uptime).
     pub fn system_info(&self) -> SystemInfo {
         self.inner.system_info()

src/runtime.rs

@@ -598,7 +598,6 @@ impl Runtime {
         self.stats_hook = Some(hook);
     }
 
-
     /// Set the sink for non-local (remote) message delivery.
     ///
     /// The sink owns all codec/transport concerns; core only knows how to hand

tools/vastai/src/config.rs

@@ -9,6 +9,7 @@ pub const ENV_MIN_INET_UP_MBPS: &str = "PP_MIN_INET_UP_MBPS";
 pub const ENV_MIN_RELIABILITY: &str = "PP_MIN_RELIABILITY";
 pub const ENV_REQUIRE_VERIFIED: &str = "PP_REQUIRE_VERIFIED";
 pub const ENV_DROP_CHEAP_FRAC: &str = "PP_DROP_CHEAP_FRAC";
+pub const ENV_MAX_DPH_TOTAL: &str = "PP_MAX_DPH_TOTAL";
 pub const ENV_LEASE_PACE_MS: &str = "PP_LEASE_PACE_MS";
 pub const ENV_BLACKLIST_HOSTS: &str = "PP_BLACKLIST_HOSTS";
 pub const ENV_ASSUME_YES: &str = "PP_ASSUME_YES";
@@ -55,6 +56,7 @@ impl SelectionPolicy {
             .unwrap_or(0.95);
         policy.require_verified = truthy_env(ENV_REQUIRE_VERIFIED);
         policy.min_up_mbps = env_optional_positive_f64(ENV_MIN_INET_UP_MBPS);
+        policy.max_dph_total = env_optional_positive_f64(ENV_MAX_DPH_TOTAL);
         policy.drop_cheap_frac = std::env::var(ENV_DROP_CHEAP_FRAC)
             .ok()
             .and_then(|s| s.trim().parse::<f64>().ok())

tools/vastai/src/filters.rs

@@ -13,5 +13,38 @@ pub(crate) fn reachable_offers(offers: Vec<Offer>, policy: &SelectionPolicy) ->
         })
         .filter(|o| o.host_id.map_or(true, |h| !blacklist.contains(&h)))
         .filter(|o| o.verification.as_deref() != Some("deverified"))
+        .filter(|o| policy.max_dph_total.is_none_or(|max| o.dph_total <= max))
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn offer(id: u64, dph_total: f64) -> Offer {
+        Offer {
+            id,
+            gpu_name: "Tesla T4".to_owned(),
+            dph_total,
+            gpu_ram: Some(16_000.0),
+            geolocation: Some("US".to_owned()),
+            inet_down_cost_per_tb: 0.0,
+            inet_up_cost_per_tb: 0.0,
+            host_id: Some(id),
+            verification: Some("unverified".to_owned()),
+        }
+    }
+
+    #[test]
+    fn max_price_cap_keeps_only_affordable_reachable_offers() {
+        let policy = SelectionPolicy {
+            max_dph_total: Some(0.10),
+            ..SelectionPolicy::default()
+        };
+
+        let reachable = reachable_offers(vec![offer(1, 0.09), offer(2, 0.11)], &policy);
+
+        assert_eq!(reachable.len(), 1);
+        assert_eq!(reachable[0].id, 1);
+    }
+}

tools/vastai/src/search.rs

@@ -47,6 +47,9 @@ pub async fn select_offer_pool_with_policy(
     if let Some(min_ram) = policy.min_gpu_ram_mb {
         query["gpu_ram"] = serde_json::json!({"gte": min_ram});
     }
+    if let Some(max_dph_total) = policy.max_dph_total {
+        query["dph_total"] = serde_json::json!({"lte": max_dph_total});
+    }
     if let Some(gpu_name) = policy.gpu_name.as_deref().filter(|s| !s.is_empty()) {
         query["gpu_name"] = serde_json::json!({"eq": gpu_name});
     }
@@ -78,14 +81,19 @@ pub async fn select_offer_pool_with_policy(
     let pool = rank_survivors(reachable, &cost, policy.drop_cheap_frac);
 
     if pool.is_empty() {
-        return Err(
+        let cap = policy
-            "no offers available (after quality/geo/host-blacklist filters and cheap-tail drop)"
+            .max_dph_total
-                .to_string(),
+            .map_or_else(|| "uncapped".to_owned(), |max| format!("max ${max:.3}/hr"));
-        );
+        return Err(format!(
+            "no offers available ({cap}, after quality/geo/host-blacklist filters and cheap-tail drop)"
+        ));
     }
+    let cap = policy
+        .max_dph_total
+        .map_or_else(|| "uncapped".to_owned(), |max| format!("max ${max:.3}/hr"));
     eprintln!(
         "select_offer_pool: {} survivor(s) for {target_count} instance(s) after \
-         per-model {:.0}% cheap-drop (cheapest ${:.3}/hr eff)",
+         per-model {:.0}% cheap-drop ({cap}, cheapest ${:.3}/hr eff)",
         pool.len(),
         policy.drop_cheap_frac * 100.0,
         cost.effective_price(&pool[0]),

tools/vastai/src/types.rs

@@ -62,6 +62,7 @@ pub struct SelectionPolicy {
     pub require_verified: bool,
     pub min_down_mbps: f64,
     pub min_up_mbps: Option<f64>,
+    pub max_dph_total: Option<f64>,
     pub blacklist_hosts: Vec<u64>,
     pub drop_cheap_frac: f64,
     pub image_size_gb: Option<f64>,
@@ -76,6 +77,7 @@ impl Default for SelectionPolicy {
             require_verified: false,
             min_down_mbps: 100.0,
             min_up_mbps: None,
+            max_dph_total: None,
             blacklist_hosts: vec![59017],
             drop_cheap_frac: 0.30,
             image_size_gb: None,

xtask/Cargo.toml

@@ -5,6 +5,7 @@ edition = "2024"
 
 [dependencies]
 serde_json = "1"
+blake3 = "1"
 
 [target.'cfg(target_os = "linux")'.dependencies]
 libc = "0.2"