Tensor and model files onto the slot API (#1665 arm D B5)

changes/internal/1665-arm-d-b5.md

@@ -0,0 +1 @@
+- Migrate tensor and model container reads and writes to the slot API.

src/model_infer.c

@@ -555,7 +555,8 @@ Value* builtin_eigen_generate(Value *arg) {
         fprintf(stderr, "eigen_generate: requires [prompt_ids, temperature, max_tokens]\n");
         TRACE_NONDET_RECORD("eigen_generate", make_list(0));
     }
-    Value *prompt_list = list_get_borrow(arg, 0);
+    EIGS_VIEW(prompt_list_view);
+    Value *prompt_list = list_get_view(arg, 0, &prompt_list_view);
     if (prompt_list->type != VAL_LIST) {
         fprintf(stderr, "eigen_generate: prompt_ids must be a list\n");
         TRACE_NONDET_RECORD("eigen_generate", make_list(0));
@@ -569,9 +570,9 @@ Value* builtin_eigen_generate(Value *arg) {
     double top_p = 0.0;
     /* #1637: optional numbers -- null keeps the default, a bool raises. */
     double mt = max_tokens;
-    eigs_opt_num(list_get_borrow(arg, 1), &temperature, __func__);
-    eigs_opt_num(list_get_borrow(arg, 2), &mt, __func__);
-    if (arg->data.list.count >= 4) eigs_opt_num(list_get_borrow(arg, 3), &top_p, __func__);
+    eigs_slot_opt_num(list_slot(arg, 1), &temperature, __func__);
+    eigs_slot_opt_num(list_slot(arg, 2), &mt, __func__);
+    if (arg->data.list.count >= 4) eigs_slot_opt_num(list_slot(arg, 3), &top_p, __func__);
     if (g_has_error) return make_null();
     max_tokens = (int)mt;
 
@@ -607,8 +608,7 @@ Value* builtin_eigen_generate(Value *arg) {
 
     int *prompt_ids = xcalloc(prompt_len, sizeof(int));
     for (int i = 0; i < prompt_len; i++) {
-        Value *v = list_get_borrow(prompt_list, i);
-        prompt_ids[i] = (int)eigs_num_arg(v, "eigen_generate");   /* #1637: a bool raises */
+        prompt_ids[i] = (int)eigs_slot_num(list_slot(prompt_list, i), "eigen_generate");   /* #1637: a bool raises */
     }
     if (g_has_error) { free(prompt_ids); return make_null(); }
 
@@ -651,12 +651,13 @@ Value* builtin_eigen_eval_loss(Value *arg) {
          * a real score. */
         ARG_GUARD(1, "eigen_eval_loss", "[prompt_ids, target_id]", make_num(-1.0));
     }
-    Value *prompt_list = list_get_borrow(arg, 0);
+    EIGS_VIEW(prompt_list_view);
+    Value *prompt_list = list_get_view(arg, 0, &prompt_list_view);
     if (prompt_list->type != VAL_LIST) {
         fprintf(stderr, "eigen_eval_loss: prompt_ids must be a list\n");
         ARG_GUARD(1, "eigen_eval_loss", "a list of prompt ids", make_num(-1.0));
     }
-    if (list_get_borrow(arg, 1)->type != VAL_NUM) {
+    if (list_elem_type(arg, 1) != VAL_NUM) {
         fprintf(stderr, "eigen_eval_loss: target_id must be a number\n");
         ARG_GUARD(1, "eigen_eval_loss", "a numeric target_id", make_num(-1.0));
     }
@@ -692,8 +693,7 @@ Value* builtin_eigen_eval_loss(Value *arg) {
 
     int *prompt_ids = xcalloc(prompt_len, sizeof(int));
     for (int i = 0; i < prompt_len; i++) {
-        Value *v = list_get_borrow(prompt_list, i);
-        int t = (int)eigs_num_arg(v, "eigen_eval_loss");          /* #1637: a bool raises */
+        int t = (int)eigs_slot_num(list_slot(prompt_list, i), "eigen_eval_loss");          /* #1637: a bool raises */
         if (t < 0 || t >= vocab_size) t = 0;
         prompt_ids[i] = t;
     }

src/model_io.c

@@ -600,12 +600,15 @@ Value* eigs_json_parse_value(const char *s, int *pos);
 
 Value* json_obj_get(Value *obj, const char *key) {
     if (!obj) return NULL;
-    if (obj->type == VAL_DICT) return dict_get(obj, key);
+    if (obj->type == VAL_DICT) {
+        int i = env_hash_find_dict(obj, key, env_hash_name(key));
+        return i < 0 ? NULL : dict_value_get_return(obj, i);
+    }
     if (obj->type != VAL_LIST) return NULL;
     for (int i = 0; i + 1 < obj->data.list.count; i += 2) {
-        Value *k = list_get_borrow(obj, i);
+        Value *k = list_get_ref(obj, i);
         if (k && k->type == VAL_STR && strcmp(k->data.str, key) == 0) {
-            return list_get_borrow(obj, i + 1);
+            return list_get_return(obj, i + 1);
         }
     }
     return NULL;

src/model_train.c

@@ -1444,16 +1444,19 @@ Value* builtin_native_train_step(Value *arg) {
         return make_str("{\"status\": \"error\", \"error\": \"Model not loaded\"}");
     }
 
-    Value *in_list = list_get_borrow(arg, 0);
-    Value *out_list = list_get_borrow(arg, 1);
+    EIGS_VIEW(in_list_view);
+
+    Value *in_list = list_get_view(arg, 0, &in_list_view);
+    EIGS_VIEW(out_list_view);
+    Value *out_list = list_get_view(arg, 1, &out_list_view);
     if (in_list->type != VAL_LIST || out_list->type != VAL_LIST) {
         return make_str("{\"status\": \"error\", \"error\": \"input_ids and output_ids must be lists\"}");
     }
 
     float lr = 0.001f;
-    if (list_get_borrow(arg, 2)->type == VAL_BOOL) { eigs_num_arg(list_get_borrow(arg, 2), __func__); return make_null(); }  /* #1637 */
-    if (list_get_borrow(arg, 2)->type == VAL_NUM) lr = eigs_list_num(arg, 2, __func__);
-    else if (list_get_borrow(arg, 2)->type == VAL_STR) lr = strtod(list_get_borrow(arg, 2)->data.str, NULL);
+    if (list_elem_type(arg, 2) == VAL_BOOL) { eigs_slot_num(list_slot(arg, 2), __func__); return make_null(); }  /* #1637 */
+    if (list_elem_type(arg, 2) == VAL_NUM) lr = eigs_list_num(arg, 2, __func__);
+    else if (list_elem_type(arg, 2) == VAL_STR) lr = strtod(list_get_ref(arg, 2)->data.str, NULL);
     if (lr <= 0 || lr > 1) lr = 0.001f;
 
     int input_len = in_list->data.list.count;
@@ -1468,12 +1471,10 @@ Value* builtin_native_train_step(Value *arg) {
     int *input_ids = xcalloc(input_len > 0 ? input_len : 1, sizeof(int));
     int *output_ids = xcalloc(output_len > 0 ? output_len : 1, sizeof(int));
     for (int i = 0; i < input_len; i++) {
-        Value *v = list_get_borrow(in_list, i);
-        input_ids[i] = (int)eigs_num_arg(v, "native_train_step");   /* #1637: a bool raises */
+        input_ids[i] = (int)eigs_slot_num(list_slot(in_list, i), "native_train_step");   /* #1637: a bool raises */
     }
     for (int i = 0; i < output_len; i++) {
-        Value *v = list_get_borrow(out_list, i);
-        output_ids[i] = (int)eigs_num_arg(v, "native_train_step");
+        output_ids[i] = (int)eigs_slot_num(list_slot(out_list, i), "native_train_step");
     }
     if (g_has_error) { free(input_ids); free(output_ids); return make_null(); }
 

tests/test_slot_escape_regress.eigs

@@ -31,4 +31,14 @@ for i in range of 70:
     deep is [deep]
 assert_eq of [deep == deep, true, "a deep nest equals itself"]
 
+# A ragged tensor holding a bare number where a row belongs: B5 read it with
+# list_get_ref (NULL for a number) and crashed; the type must be tested first.
+assert_eq of [add of [[[1, 2], 3], [1, 1]], [[2, 3], [4, 4]], "ragged broadcast add treats a number row as a scalar"]
+assert_eq of [gather of [[[1, 2], 5], [0, 0]], [1, 0], "gather on a ragged tensor"]
+try:
+    r is relu of ([[7, 1], 1])
+    assert_eq of ["no raise", "raise", "relu on a ragged tensor raises"]
+catch e:
+    assert_eq of [e.kind, "type_mismatch", "relu on a ragged tensor raises a type error"]
+
 test_summary of null

tools/container_migrated_files.txt

@@ -4,3 +4,7 @@ src/eigenscript.c
 src/trace.c
 src/vm.c
 src/builtins.c
+src/builtins_tensor.c
+src/model_infer.c
+src/model_train.c
+src/model_io.c

tools/num_read_allowlist.txt

@@ -5,7 +5,6 @@
 # count), when a function with raw tokens is missing here, or a row is stale.
 # Builtins never belong here: they read through eigs_num_arg / eigs_list_num /
 # eigs_opt_num / eigs_elem_num, which raise on a bool.
-builtins_tensor.c|builtin_numerical_grad_rows|3|in-place writes of a matrix cell that tensor_cells_numeric proved a number before any mutation
 builtins_tensor.c|builtin_tensor_matmul|1|writes the NaN-boxed null sentinel into a result buffer cell (EIGS_STRICT=0 path); a write, no Value is read
 container.h|dict_value_get_num|1|accessor numeric fast path: reads the number only behind its own type == VAL_NUM test (#1665 arm D)
 container.h|dict_value_get_owned|1|owned numeric read copies a proven VAL_NUM box for flip-safe ownership

tools/strict_shape_contracts.json

@@ -1455,7 +1455,7 @@
         "b38ee26c123f3a682f61cfacc3cd75d7dd6423aa2eafbceb49e64b2e63ee47bf"
       ],
       "call_eigs_fn": [
-        "123bbc53a201557a82d9c75caf6dc231d9c2485e9d2ccce3a296ccca5c9f95d6"
+        "7b0b1ef9ee65309c4c965af3c4167aa1e832da6185c80f16f3b05bb42d2b860c"
       ],
       "env_set_local": [
         "bde886fc51cd3b927e7c8c487a7a0d270453133c4e598506927a2d549788afbf"
@@ -2268,7 +2268,7 @@
         "f6bb04ca6134dd41d1a67ee91ea2b9587e13477f13ed51afaa3834de1d4f2b78"
       ],
       "tensor_unary": [
-        "4abceb6bf32d03e72bf0b7245447987b78885d079328ae6504e48bf8afa170f9"
+        "6cd5986e901ac953b121ba49c8b8d0eac1c53a1cf27cb8889b7f8f9a7e1b9a11"
       ]
     }
   },
@@ -2309,7 +2309,7 @@
         "40100cc6cf4d92300cf1d3fbae23cb21adf4241bd01edecc9a683a07fb30bc01"
       ],
       "tensor_to_flat": [
-        "935bb2e66301b7e41ac4d999bb1e5bfee0b4b4ab8afd54bb73d0369f71c67898"
+        "d6c74fa5dc6c5d4d702b6f2eb1746dbe61271fd05f9bae972bdefa05d607cedc"
       ]
     }
   },
@@ -2325,11 +2325,11 @@
     ],
     "disposition": "exempt",
     "body_sha256": [
-      "6923863238d3db9fb71ddeb0c0b56d92e9436253f886350d6b89560b5f55e963"
+      "a247feeafdfc05d844eb1bdd0f3eed2be3cf37a4801ca34add1628d8f141b8a5"
     ],
     "source_sha256": {
       "builtin_tensor_load": [
-        "6923863238d3db9fb71ddeb0c0b56d92e9436253f886350d6b89560b5f55e963"
+        "a247feeafdfc05d844eb1bdd0f3eed2be3cf37a4801ca34add1628d8f141b8a5"
       ]
     }
   },
@@ -2352,7 +2352,7 @@
         "ff086ad360320823ab29cb9cfbeed6ec340ae78f959cb18c187c227db55f947f"
       ],
       "tensor_unary": [
-        "4abceb6bf32d03e72bf0b7245447987b78885d079328ae6504e48bf8afa170f9"
+        "6cd5986e901ac953b121ba49c8b8d0eac1c53a1cf27cb8889b7f8f9a7e1b9a11"
       ]
     }
   },
@@ -2368,11 +2368,11 @@
     ],
     "disposition": "exempt",
     "body_sha256": [
-      "28d209db0561173b9426149c963f32bbb8a1d4bd15ded1aad450e3e84e4b986e"
+      "868db4f26fa6b315a63c25e1474c2ba2bbea74c1ff294876bfef6c715906a0e3"
     ],
     "source_sha256": {
       "builtin_tensor_log_softmax": [
-        "28d209db0561173b9426149c963f32bbb8a1d4bd15ded1aad450e3e84e4b986e"
+        "868db4f26fa6b315a63c25e1474c2ba2bbea74c1ff294876bfef6c715906a0e3"
       ]
     }
   },
@@ -2431,10 +2431,10 @@
         "f6eec99bd0153900403028372a294cf84be9485ce98b7ad8843ce4e2a5d390ee"
       ],
       "tensor_flatten_recursive": [
-        "3844bd92ee4238bec0b3bd2645eb4bd17d18a6d681a7ad7a8e21291282e33d14"
+        "be389e2e858e13f1dc09f309722e7e6ccebf4bb2bb829b2b1ad176401d2e86ae"
       ],
       "tensor_total": [
-        "a0a594c0e05d41ef7ef4a8c262d56646d628f3b3f5c4f94b40e9cf1637d8ca03"
+        "e08db1676199c3179be8948bba78f43af63fecc404ef75b8d0b55ac59be72484"
       ]
     }
   },
@@ -2469,7 +2469,7 @@
         "a345f0c30c4b1b80943b9c718c0002fc19981d0336a9a1ea0aca3b5966bf8462"
       ],
       "tensor_unary": [
-        "4abceb6bf32d03e72bf0b7245447987b78885d079328ae6504e48bf8afa170f9"
+        "6cd5986e901ac953b121ba49c8b8d0eac1c53a1cf27cb8889b7f8f9a7e1b9a11"
       ]
     }
   },
@@ -2492,10 +2492,10 @@
         "817e544211920e1216ff102393d66f3cb7a8def4a9ea05611872433b5ad8f87c"
       ],
       "tensor_flatten_recursive": [
-        "3844bd92ee4238bec0b3bd2645eb4bd17d18a6d681a7ad7a8e21291282e33d14"
+        "be389e2e858e13f1dc09f309722e7e6ccebf4bb2bb829b2b1ad176401d2e86ae"
       ],
       "tensor_total": [
-        "a0a594c0e05d41ef7ef4a8c262d56646d628f3b3f5c4f94b40e9cf1637d8ca03"
+        "e08db1676199c3179be8948bba78f43af63fecc404ef75b8d0b55ac59be72484"
       ]
     }
   },
@@ -2533,7 +2533,7 @@
         "9829b048a1af1233bfeae6329573e79d83cd7acb01e548d3737011a625fbdc02"
       ],
       "tensor_to_flat": [
-        "935bb2e66301b7e41ac4d999bb1e5bfee0b4b4ab8afd54bb73d0369f71c67898"
+        "d6c74fa5dc6c5d4d702b6f2eb1746dbe61271fd05f9bae972bdefa05d607cedc"
       ]
     }
   },
@@ -2573,11 +2573,11 @@
     ],
     "disposition": "exempt",
     "body_sha256": [
-      "d5ab1a467607560bfce18d36abd18fa5b25d7fe43179da3c305e29f379b027e7"
+      "f78ba93eb0b22a35b1c8a328a1f557b7096bf05bdbf38fe3259876cc7d1b8bb9"
     ],
     "source_sha256": {
       "builtin_tensor_shape": [
-        "d5ab1a467607560bfce18d36abd18fa5b25d7fe43179da3c305e29f379b027e7"
+        "f78ba93eb0b22a35b1c8a328a1f557b7096bf05bdbf38fe3259876cc7d1b8bb9"
       ]
     }
   },
@@ -2609,7 +2609,7 @@
         "40100cc6cf4d92300cf1d3fbae23cb21adf4241bd01edecc9a683a07fb30bc01"
       ],
       "tensor_to_flat": [
-        "935bb2e66301b7e41ac4d999bb1e5bfee0b4b4ab8afd54bb73d0369f71c67898"
+        "d6c74fa5dc6c5d4d702b6f2eb1746dbe61271fd05f9bae972bdefa05d607cedc"
       ]
     }
   },
@@ -2632,7 +2632,7 @@
         "f38d9057e3c1e2a049e7cd4d4f38df6aace72af77d1a718b357b4f2a5bf4f54d"
       ],
       "tensor_unary": [
-        "4abceb6bf32d03e72bf0b7245447987b78885d079328ae6504e48bf8afa170f9"
+        "6cd5986e901ac953b121ba49c8b8d0eac1c53a1cf27cb8889b7f8f9a7e1b9a11"
       ]
     }
   },
@@ -2667,10 +2667,10 @@
         "196480c8b98743ea57d8bf3bc4bdf0b0fdc3ca70c56ef6cd1289342e15fcd544"
       ],
       "tensor_flatten_recursive": [
-        "3844bd92ee4238bec0b3bd2645eb4bd17d18a6d681a7ad7a8e21291282e33d14"
+        "be389e2e858e13f1dc09f309722e7e6ccebf4bb2bb829b2b1ad176401d2e86ae"
       ],
       "tensor_total": [
-        "a0a594c0e05d41ef7ef4a8c262d56646d628f3b3f5c4f94b40e9cf1637d8ca03"
+        "e08db1676199c3179be8948bba78f43af63fecc404ef75b8d0b55ac59be72484"
       ]
     }
   },
@@ -2698,11 +2698,11 @@
     ],
     "disposition": "exempt",
     "body_sha256": [
-      "aad531cea0cb998c5ab4bcfbc19ca755f163227267af03b6ad6f3ae66a8c751f"
+      "0de72aaa352907156e4244579f587cadc755557642f4e297485a5c4a019233eb"
     ],
     "source_sha256": {
       "builtin_tensor_zeros_like": [
-        "aad531cea0cb998c5ab4bcfbc19ca755f163227267af03b6ad6f3ae66a8c751f"
+        "0de72aaa352907156e4244579f587cadc755557642f4e297485a5c4a019233eb"
       ],
       "make_buffer_like": [
         "40100cc6cf4d92300cf1d3fbae23cb21adf4241bd01edecc9a683a07fb30bc01"