feat(diagnose): answer whether a model will run before downloading it

README.md

@@ -227,6 +227,7 @@ form works depends on the engine your GPU selects.
 | `rocm` | Open the launcher menu (setup, serve, diagnose, chat, dashboard) |
 | `rocm examine` | Check GPU, ROCm install, engines, and managed folders |
 | `rocm diagnose` | Match this machine against known ROCm/PyTorch/llama.cpp failure modes |
+| `rocm diagnose --model <model>` | Say whether a model will run here, before downloading it |
 | `rocm fix [<fix-id>]` | Apply a fix reported by `rocm diagnose` |
 | `rocm install sdk` | Install TheRock ROCm wheels into a managed Python environment |
 | `rocm install driver` | Install the AMD kernel driver on Linux |
@@ -261,6 +262,7 @@ the JSON report, not the human-readable one.
 
 ```
 rocm diagnose [--symptom TEXT] [--top N] [--json] [--distro [NAME]]
+rocm diagnose --model <model> [--json]
 rocm fix [<fix-id>] [--yes] [--dry-run] [--device-index N]
 ```
 
@@ -452,6 +454,27 @@ rocm engines shell <engine>   [--runtime-id KEY | --env-id ID] [--shell PATH]
 
 Supported engines: `lemonade`, `vllm`.
 
+### Will a model run here?
+
+Ask before downloading anything:
+
+```
+rocm diagnose --model <model> [--json]
+```
+
+Answers in seconds, from the curated recipe and this machine's GPU — it
+fetches no weights and makes no network call. The verdict is `ready`,
+`degraded`, `blocked`, or `undetermined`. A `ready` answer also names the
+engine `rocm serve` would use; a `blocked` one names curated models that would
+run here instead.
+
+`undetermined` is a real answer and not a failure: it is what you get when the
+recipe catalog could not be read, when this machine's GPU memory could not be
+measured, or when the model is not one of the curated recipes (`rocm model`
+lists those). None of those say anything about whether the model fits, so none
+of them are reported as though they did — `rocm serve` still accepts a model
+outside the catalog, this just cannot tell you in advance how it will go.
+
 ### Model serving
 
 Start a local OpenAI-compatible model server:

apps/rocm/src/remote/doctor.rs

@@ -338,6 +338,11 @@ mod tests {
                 url: "https://example.invalid/issues".to_owned(),
             },
             out_of_scope: None,
+            // These fixtures exercise the remote renderer, which prints the
+            // environment half of a report. A model verdict rides on top of
+            // that half rather than replacing it, so there is nothing for these
+            // to say about one.
+            model: None,
         }
     }
 

crates/rocm-core/src/diagnose.rs

@@ -87,6 +87,16 @@ pub struct DiagnoseReport {
     /// avoid emitting bare-metal-Linux diagnoses that don't apply.
     #[serde(default)]
     pub out_of_scope: Option<String>,
+    /// The verdict on a model, when `--model` named one.
+    ///
+    /// Attached by the caller after the fact rather than produced by
+    /// [`diagnose`]: answering it needs the host's GPU memory and the engine
+    /// `serve` would select, neither of which an [`Examination`] carries. It
+    /// rides on this report rather than replacing it because the environment
+    /// answer stays true and useful either way — a blocked model on a host whose
+    /// driver is also misconfigured is two findings, not one.
+    #[serde(default)]
+    pub model: Option<crate::model_readiness::ModelReadiness>,
 }
 
 /// Whether any diagnosis cleared [`MIN_SCORE_FOR_MATCH`].
@@ -2115,6 +2125,20 @@ fn catalog_covers(e: &Examination) -> bool {
         .any(|(_, applicable)| applicable.contains(&family))
 }
 
+/// Where to report something this CLI could not answer.
+///
+/// The one place a `Route` is built, so every command that has to say "I don't
+/// recognise this" sends the user to the same tracker for the same target —
+/// `rocm diagnose --model` has the same problem for a model the catalog does not
+/// carry as `rocm diagnose` has for a symptom it does not recognise.
+#[must_use]
+pub fn upstream_route(target: &str) -> Route {
+    Route {
+        target: target.to_owned(),
+        url: upstream_tracker(target).to_owned(),
+    }
+}
+
 /// Where to send a user when nothing in the catalog matched.
 ///
 /// Keyed off the *host-detected* framework, which `Examination::probe` only
@@ -2132,10 +2156,7 @@ fn route_when_no_match(e: &Examination) -> Route {
         "llama-cpp" => "llama-cpp",
         _ => "rocm-core",
     };
-    Route {
-        target: target.to_owned(),
-        url: upstream_tracker(target).to_owned(),
-    }
+    upstream_route(target)
 }
 
 /// Diagnose an examination against the closed catalog.
@@ -2162,6 +2183,7 @@ pub fn diagnose(e: &Examination, symptom: &str) -> DiagnoseReport {
         high_confidence_threshold: HIGH_CONFIDENCE,
         route_when_no_match: route_when_no_match(e),
         out_of_scope,
+        model: None,
     }
 }
 

crates/rocm-core/src/lib.rs

@@ -31,14 +31,15 @@ pub mod diagnose;
 pub mod disk_space;
 pub mod examine;
 pub mod fix;
+pub mod model_readiness;
 pub mod openmpi;
 pub mod proc_lifecycle;
 pub mod runtime;
 #[cfg(test)]
 mod test_env;
 pub mod uv;
 pub use diagnose::{
-    DiagnoseReport, Diagnosis, Fix, diagnose as run_diagnose,
+    DiagnoseReport, Diagnosis, Fix, Route, diagnose as run_diagnose,
     render_report_text as render_diagnose_text,
 };
 pub use disk_space::{